File size: 124,606 Bytes
894fa47 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 | # This Python 3 environment comes with many helpful analytics libraries installed
# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python
# For example, here's several helpful packages to load
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
# Input data files are available in the read-only "../input/" directory
# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory
import os
for dirname, _, filenames in os.walk('/kaggle/input'):
for filename in filenames:
print(os.path.join(dirname, filename))
# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using "Save & Run All"
# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session
import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import cv2
import os
import pandas as pd
from scipy.ndimage import gaussian_filter
# -------------------------------------------
# Configuration
# -------------------------------------------
OUTPUT_DIR = "/kaggle/working"
GRADCAM_OUTPUT_DIR = os.path.join(OUTPUT_DIR, 'gradcam_plus_plus_results')
os.makedirs(GRADCAM_OUTPUT_DIR, exist_ok=True)
print("="*80)
print("HIERARCHICAL TRI-HEAD GRAD-CAM++ CONFIGURATION")
print("="*80)
print(f"Output directory: {GRADCAM_OUTPUT_DIR}")
print("="*80)
# -------------------------------------------
# Class Mappings
# -------------------------------------------
DISEASE_CLASS_MAPPING = {
0: "Breast_cancer",
1: "annrbc-anemia_processed",
2: "colon_processed",
3: "leukemia_processed",
4: "lung_processed",
5: "oral-cancer_processed",
6: "ovarian-cancer_processed",
7: "sickle-cell-new_processed",
8: "thalassemia_processed",
}
SEVERITY_CLASS_MAPPING = {
0: "Normal",
1: "Abnormal",
}
# -------------------------------------------
# Grad-CAM++ Implementation for ViT
# -------------------------------------------
# -------------------------------------------
# Grad-CAM++ Implementation for ViT (FIXED for tuple outputs)
# -------------------------------------------
class GradCAM:
"""
Standard Grad-CAM implementation for Vision Transformers
Simplified - no second-order gradients, just straightforward CAM
"""
def __init__(self, model, target_layer):
"""
Args:
model: Your Phase3 hierarchical model
target_layer: The layer to hook (typically last transformer block)
"""
self.model = model
self.target_layer = target_layer
self.gradients = None
self.activations = None
# Register hooks
self.handlers = []
self._register_hooks()
def _register_hooks(self):
"""Register forward and backward hooks on target layer"""
def forward_hook(module, input, output):
# Handle tuple output (DINOv2 returns tuple)
if isinstance(output, tuple):
self.activations = output[0].detach()
print(f" 🪝 Forward hook: Captured from tuple, shape {output[0].shape}")
else:
self.activations = output.detach()
print(f" 🪝 Forward hook: Captured tensor, shape {output.shape}")
def backward_hook(module, grad_input, grad_output):
# Handle tuple output in gradients
if isinstance(grad_output, tuple):
grad = grad_output[0]
if grad is not None:
self.gradients = grad.detach()
print(f" 🪝 Backward hook: Captured from tuple, shape {grad.shape}")
else:
if grad_output is not None:
self.gradients = grad_output.detach()
print(f" 🪝 Backward hook: Captured tensor, shape {grad_output.shape}")
# Register hooks
self.handlers.append(
self.target_layer.register_forward_hook(forward_hook)
)
self.handlers.append(
self.target_layer.register_full_backward_hook(backward_hook)
)
def remove_hooks(self):
"""Remove all hooks"""
for handle in self.handlers:
handle.remove()
def generate_cam(self, class_idx, logits):
"""
Generate standard Grad-CAM heatmap
Args:
class_idx: Target class index
logits: Model output logits
Returns:
cam: Grad-CAM heatmap (H, W)
"""
# Zero gradients
self.model.zero_grad()
# Backward pass
one_hot = torch.zeros_like(logits)
one_hot[0, class_idx] = 1
logits.backward(gradient=one_hot, retain_graph=True)
# ========================================
# 🔍 GRADIENT FLOW DEBUGGING
# ========================================
print(f"\n 🔍 GRADIENT FLOW CHECK:")
print(f" {'='*60}")
# Check if gradients were captured
if self.gradients is None:
print(f" ❌ CRITICAL: No gradients captured!")
return np.zeros((14, 14))
else:
print(f" ✅ Gradients captured: {self.gradients.shape}")
print(f" Min: {self.gradients.min().item():.6f}, Max: {self.gradients.max().item():.6f}")
print(f" Mean: {self.gradients.mean().item():.6f}, Std: {self.gradients.std().item():.6f}")
# Check if activations were captured
if self.activations is None:
print(f" ❌ CRITICAL: No activations captured!")
return np.zeros((14, 14))
else:
print(f" ✅ Activations captured: {self.activations.shape}")
print(f" Min: {self.activations.min().item():.6f}, Max: {self.activations.max().item():.6f}")
print(f" {'='*60}\n")
# ========================================
# Standard Grad-CAM computation
# gradients: [B, N, D]
# activations: [B, N, D]
# Step 1: Global average pooling on gradients to get weights
# Take mean across spatial dimension (tokens) for each channel
weights = self.gradients.mean(dim=1, keepdim=True) # [B, 1, D]
print(f" 📊 Weights (channel importance):")
print(f" Shape: {weights.shape}")
print(f" Min: {weights.min().item():.6f}, Max: {weights.max().item():.6f}")
# Step 2: Weighted combination of activation maps
# weights: [B, 1, D]
# activations: [B, N, D]
# Result: [B, N] - one value per token
cam = (weights * self.activations).sum(dim=2) # Sum across channels (D)
print(f" 📊 CAM before ReLU:")
print(f" Shape: {cam.shape}")
print(f" Min: {cam.min().item():.6f}, Max: {cam.max().item():.6f}")
# Step 3: Apply ReLU (only keep positive contributions)
#cam = F.relu(cam)
print(f" 📊 CAM after ReLU:")
print(f" Min: {cam.min().item():.6f}, Max: {cam.max().item():.6f}")
# Step 4: Remove batch dimension
cam = cam[0] # [N]
# Step 5: Remove CLS token (first token in ViT)
if cam.shape[0] > 1:
cam = cam[1:]
print(f" 🎯 Removed CLS token, remaining tokens: {cam.shape[0]}")
# Step 6: Reshape to spatial grid
grid_size = int(np.sqrt(cam.shape[0]))
print(f" 📐 Grid size: {grid_size}x{grid_size}")
cam = cam.reshape(grid_size, grid_size)
# Step 7: Normalize to [0, 1]
cam_min = cam.min()
cam_max = cam.max()
if cam_max > cam_min:
cam = (cam - cam_min) / (cam_max - cam_min)
print(f" ✅ Normalized CAM: min={cam.min().item():.4f}, max={cam.max().item():.4f}")
else:
print(f" ⚠️ WARNING: No variation in CAM (all same value)")
cam = torch.zeros_like(cam)
print(f" 📊 Final CAM mean: {cam.mean().item():.6f}\n")
return cam.cpu().numpy()
class GradCAMVisionTransformer:
"""
Wrapper to apply standard Grad-CAM to Vision Transformer models
"""
def __init__(self, model, device):
self.model = model
self.device = device
self.model.eval()
# Find the last transformer block
self.target_layer = self._find_target_layer()
print(f" 🎯 Target layer for Grad-CAM: {self.target_layer}")
def _find_target_layer(self):
"""
Find the last transformer block in DINOv2/Phikon model
"""
print("\n🔍 Searching for target layer in DINOv2 architecture...")
try:
if hasattr(self.model, 'backbone'):
vit_model = self.model.backbone.vit
print(f"✅ Found backbone.vit: {type(vit_model).__name__}")
else:
raise AttributeError("No backbone found")
if hasattr(vit_model, 'encoder') and hasattr(vit_model.encoder, 'layer'):
num_layers = len(vit_model.encoder.layer)
last_layer = vit_model.encoder.layer[-1]
print(f"✅ Found encoder with {num_layers} layers")
print(f"✅ Target layer: encoder.layer[-1] (layer {num_layers-1})")
return last_layer
else:
raise AttributeError("No encoder.layer found")
except AttributeError as e:
print(f"❌ Error: {e}")
raise ValueError("Could not find DINOv2 encoder layers")
def generate_heatmap(self, image_tensor, target_class_idx, head_type='disease'):
"""
Generate Grad-CAM heatmap for a specific head
Args:
image_tensor: Input image tensor [1, 3, H, W]
target_class_idx: Target class index
head_type: 'disease', 'severity', or 'stage'
Returns:
heatmap: Grad-CAM heatmap resized to input image size
"""
# Create Grad-CAM instance
gradcam = GradCAM(self.model, self.target_layer)
# Forward pass with gradients enabled
image_tensor = image_tensor.to(self.device)
image_tensor.requires_grad = True
# Get logits based on head type
disease_logits, severity_logits, stage_logits, _ = self.model([image_tensor], enable_gradients=True)
if head_type == 'disease':
logits = disease_logits
elif head_type == 'severity':
disease_pred_idx = disease_logits.argmax(dim=1).item()
disease_name = DISEASE_CLASS_MAPPING.get(disease_pred_idx, f"Unknown_{disease_pred_idx}")
logits = severity_logits[disease_name]
elif head_type == 'stage':
logits = stage_logits
else:
raise ValueError(f"Unknown head_type: {head_type}")
# Generate CAM
cam = gradcam.generate_cam(target_class_idx, logits)
# Clean up hooks
gradcam.remove_hooks()
# Resize to match input image size
H, W = image_tensor.shape[2], image_tensor.shape[3]
cam_resized = cv2.resize(cam, (W, H), interpolation=cv2.INTER_CUBIC)
# Smooth the heatmap
cam_smooth = gaussian_filter(cam_resized, sigma=2)
return cam_smooth
# -------------------------------------------
# Helper Functions
# -------------------------------------------
def create_gradcam_overlay(image_array, heatmap, alpha=0.5, colormap='jet'):
"""
Create a visual overlay of Grad-CAM++ heatmap on original image
Uses red-yellow colormap like traditional Grad-CAM
Args:
image_array: Original image as numpy array (H, W, 3)
heatmap: Grad-CAM++ heatmap (H, W), values in [0, 1]
alpha: Transparency of heatmap overlay
colormap: Matplotlib colormap name
Returns:
Overlayed image as numpy array (H, W, 3) in range [0, 1]
"""
# Normalize image to [0, 1]
img_normalized = image_array.astype(np.float32) / 255.0
# Resize heatmap to match image size if needed
target_h, target_w = img_normalized.shape[:2]
if heatmap.shape != (target_h, target_w):
print(f" 📐 Resizing heatmap from {heatmap.shape} to ({target_h}, {target_w})")
heatmap = cv2.resize(heatmap, (target_w, target_h), interpolation=cv2.INTER_CUBIC)
# Ensure heatmap is in [0, 1]
heatmap = np.clip(heatmap, 0, 1)
# Apply colormap - Fixed for newer matplotlib
import matplotlib
cmap = matplotlib.colormaps.get_cmap(colormap)
heatmap_colored = cmap(heatmap)[:, :, :3] # Remove alpha channel
# Blend with original image
overlay = img_normalized * (1 - alpha) + heatmap_colored * alpha
overlay = np.clip(overlay, 0, 1)
return overlay
def load_and_preprocess_image(img_path, target_size=224):
"""Load image and preprocess for both visualization and model input"""
try:
if img_path.lower().endswith(('.svs', '.ndpi')):
slide = openslide.OpenSlide(img_path)
img_pil = slide.get_thumbnail((target_size, target_size))
slide.close()
elif img_path.lower().endswith('.tif'):
try:
slide = openslide.OpenSlide(img_path)
img_pil = slide.get_thumbnail((target_size, target_size))
slide.close()
except:
img_pil = Image.open(img_path).convert('RGB')
img_pil = img_pil.resize((target_size, target_size), Image.BILINEAR)
else:
img_pil = Image.open(img_path).convert('RGB')
img_pil = img_pil.resize((target_size, target_size), Image.BILINEAR)
img_array = np.array(img_pil)
img_tensor = test_transform(img_pil).unsqueeze(0)
return img_pil, img_array, img_tensor
except Exception as e:
print(f"Error loading image {img_path}: {e}")
return None, None, None
def extract_stage_label(img_path):
"""
Extract stage label using strict hierarchy rules.
Valid structures:
- .../abnormal/test/image.png -> no stage -> return -1
- .../abnormal/<stage_name>/test/image.png -> stage exists -> return <stage_name>
"""
path_parts = img_path.split(os.sep)
try:
# Find 'test' folder
test_idx = path_parts.index('test')
# Folder immediately above 'test'
candidate = path_parts[test_idx - 1]
# If abnormal is directly above test → no stage
if candidate.lower() == 'abnormal':
return -1
# Otherwise, this folder is the stage
return candidate
except (ValueError, IndexError):
# 'test' not found or malformed path
return -1
# -------------------------------------------
# Main Tri-Head Grad-CAM++ Analysis
# -------------------------------------------
def run_tri_head_gradcam_plus_plus_analysis(model, device, collected_images):
"""
Run Grad-CAM++ analysis on disease head (Level 1), severity head (Level 2), and stage head (Level 3)
Stage head is only analyzed when ground truth is abnormal and stage label is not -1
Returns comprehensive dictionaries with all analysis results
"""
print("\n" + "="*80)
print("🔍 STARTING TRI-HEAD GRAD-CAM++ ANALYSIS")
print("="*80)
if not collected_images:
print("❌ No images provided")
return []
print(f"✅ Processing {len(collected_images)} images\n")
# Initialize Grad-CAM++ wrapper
gradcam_wrapper = GradCAMVisionTransformer(model, device)
all_results = []
for idx, img_path in enumerate(collected_images):
print(f"\n[{idx+1}/{len(collected_images)}] Processing: {os.path.basename(img_path)}")
try:
# Load image
img_pil, img_array, img_tensor = load_and_preprocess_image(img_path, target_size=224)
if img_tensor is None:
print(f" ❌ Failed to load image")
continue
# Extract metadata
path_parts = img_path.split(os.sep)
dataset_name = "Unknown"
true_label = "unknown"
stage_label = extract_stage_label(img_path)
for part in path_parts:
if part in ["ovarian-cancer_processed", "oral-cancer_processed",
"Breast_cancer", "colon_processed", "lung_processed",
"annrbc-anemia_processed", "leukemia_processed",
"sickle-cell-new_processed", "thalassemia_processed"]:
dataset_name = part
if part in ["normal", "abnormal"]:
true_label = part
break
# ===== STEP 1: Get predictions from all heads =====
with torch.no_grad():
img_tensor_device = img_tensor.to(device)
disease_logits, severity_logits_dict, stage_logits, _ = model([img_tensor_device])
# Disease prediction (Level 1)
disease_pred_idx = disease_logits.argmax(dim=1).item()
disease_probs = F.softmax(disease_logits, dim=1)
disease_confidence = disease_probs[0, disease_pred_idx].item()
disease_name = DISEASE_CLASS_MAPPING.get(disease_pred_idx, f"Unknown_{disease_pred_idx}")
disease_all_probs = disease_probs[0].cpu().numpy()
print(f" 📊 Level 1 (Disease): {disease_name}")
print(f" Index: {disease_pred_idx}, Confidence: {disease_confidence:.4f}")
# Severity prediction (Level 2)
severity_logits = severity_logits_dict[disease_name]
severity_pred_idx = severity_logits.argmax(dim=1).item()
severity_probs = F.softmax(severity_logits, dim=1)
severity_confidence = severity_probs[0, severity_pred_idx].item()
severity_label_text = SEVERITY_CLASS_MAPPING.get(severity_pred_idx, f"Unknown_{severity_pred_idx}")
severity_all_probs = severity_probs[0].cpu().numpy()
print(f" 📊 Level 2 (Severity): {severity_label_text}")
print(f" Index: {severity_pred_idx}, Confidence: {severity_confidence:.4f}")
# Stage prediction (Level 3) - if available
stage_pred_idx = None
stage_confidence = None
stage_all_probs = None
if stage_logits is not None:
stage_pred_idx = stage_logits.argmax(dim=1).item()
stage_probs = F.softmax(stage_logits, dim=1)
stage_confidence = stage_probs[0, stage_pred_idx].item()
stage_all_probs = stage_probs[0].cpu().numpy()
print(f" 📊 Level 3 (Stage): Stage {stage_pred_idx}")
print(f" Confidence: {stage_confidence:.4f}")
print(f" Ground Truth Stage: {stage_label}")
# ===== STEP 2: Generate Grad-CAM++ for disease head =====
print(f"\n 🔥 Generating Grad-CAM++ for Disease Head...")
disease_heatmap = gradcam_wrapper.generate_heatmap(
img_tensor.clone(),
disease_pred_idx,
head_type='disease'
)
# Create overlay
disease_overlay = create_gradcam_overlay(img_array, disease_heatmap, alpha=0.5)
print(f" ✅ Disease heatmap generated")
print(f" Min: {disease_heatmap.min():.4f}, Max: {disease_heatmap.max():.4f}")
# ===== STEP 3: Generate Grad-CAM++ for severity head =====
print(f" 🔥 Generating Grad-CAM++ for Severity Head...")
severity_heatmap = gradcam_wrapper.generate_heatmap(
img_tensor.clone(),
severity_pred_idx,
head_type='severity'
)
# Create overlay
severity_overlay = create_gradcam_overlay(img_array, severity_heatmap, alpha=0.5)
print(f" ✅ Severity heatmap generated")
print(f" Min: {severity_heatmap.min():.4f}, Max: {severity_heatmap.max():.4f}")
# ===== STEP 4: Generate Grad-CAM++ for stage head (conditional) =====
stage_heatmap = None
stage_overlay = None
include_stage_analysis = False
# Check conditions: abnormal ground truth AND stage label != -1
if true_label == "abnormal" and stage_label != -1 and stage_logits is not None:
include_stage_analysis = True
print(f" 🔥 Generating Grad-CAM++ for Stage Head (GT: abnormal, Stage: {stage_label})...")
stage_heatmap = gradcam_wrapper.generate_heatmap(
img_tensor.clone(),
stage_pred_idx,
head_type='stage'
)
# Create overlay
stage_overlay = create_gradcam_overlay(img_array, stage_heatmap, alpha=0.5)
print(f" ✅ Stage heatmap generated")
print(f" Min: {stage_heatmap.min():.4f}, Max: {stage_heatmap.max():.4f}")
else:
reason = []
if true_label != "abnormal":
reason.append(f"true_label='{true_label}'")
if stage_label == -1:
reason.append("stage_label=-1")
if stage_logits is None:
reason.append("stage_logits=None")
print(f" ⏭️ Skipping Stage Head Analysis ({', '.join(reason)})")
# ===== STEP 5: Create Union Heatmap =====
if include_stage_analysis:
# Average of all three heatmaps
union_heatmap = (disease_heatmap + severity_heatmap + stage_heatmap) / 3.0
print(f" 📊 Union Heatmap: Average of 3 heads (Disease + Severity + Stage)")
else:
# Average of two heatmaps
union_heatmap = (disease_heatmap + severity_heatmap) / 2.0
print(f" 📊 Union Heatmap: Average of 2 heads (Disease + Severity)")
union_overlay = create_gradcam_overlay(img_array, union_heatmap, alpha=0.5)
# ===== STEP 6: Calculate statistics =====
disease_mean_activation = float(disease_heatmap.mean())
disease_max_activation = float(disease_heatmap.max())
severity_mean_activation = float(severity_heatmap.mean())
severity_max_activation = float(severity_heatmap.max())
stage_mean_activation = None
stage_max_activation = None
if stage_heatmap is not None:
stage_mean_activation = float(stage_heatmap.mean())
stage_max_activation = float(stage_heatmap.max())
union_mean_activation = float(union_heatmap.mean())
union_max_activation = float(union_heatmap.max())
# ===== STEP 7: Compile comprehensive results dictionary =====
result_dict = {
# ===== Image Information =====
'filename': os.path.basename(img_path),
'full_path': img_path,
'dataset_name': dataset_name,
'true_label': true_label,
'stage_label': stage_label,
'include_stage_analysis': include_stage_analysis,
# ===== Original Image =====
'image': img_array,
# ===== Level 1: Disease Head Results =====
'level1_disease': {
'predicted_class': disease_name,
'predicted_idx': disease_pred_idx,
'confidence': disease_confidence,
'all_probabilities': disease_all_probs,
'heatmap_raw': disease_heatmap,
'heatmap_overlay': disease_overlay,
'activation_stats': {
'mean': disease_mean_activation,
'max': disease_max_activation,
}
},
# ===== Level 2: Severity Head Results =====
'level2_severity': {
'predicted_class': severity_label_text,
'predicted_idx': severity_pred_idx,
'confidence': severity_confidence,
'all_probabilities': severity_all_probs,
'heatmap_raw': severity_heatmap,
'heatmap_overlay': severity_overlay,
'activation_stats': {
'mean': severity_mean_activation,
'max': severity_max_activation,
}
},
# ===== Level 3: Stage Head Results (conditional) =====
'level3_stage': {
'predicted_idx': stage_pred_idx,
'confidence': stage_confidence,
'all_probabilities': stage_all_probs,
'heatmap_raw': stage_heatmap,
'heatmap_overlay': stage_overlay,
'activation_stats': {
'mean': stage_mean_activation,
'max': stage_max_activation,
} if stage_heatmap is not None else None
},
# ===== Union Results =====
'union': {
'heatmap_raw': union_heatmap,
'heatmap_overlay': union_overlay,
'num_heads_averaged': 3 if include_stage_analysis else 2,
'activation_stats': {
'mean': union_mean_activation,
'max': union_max_activation,
}
},
# ===== Legacy Fields =====
'disease_heatmap': disease_heatmap,
'severity_heatmap': severity_heatmap,
'stage_heatmap': stage_heatmap,
'disease_pred': disease_name,
'disease_idx': disease_pred_idx,
'disease_conf': disease_confidence,
'severity_pred': severity_label_text,
'severity_idx': severity_pred_idx,
'severity_conf': severity_confidence,
'stage_pred_idx': stage_pred_idx,
'stage_conf': stage_confidence,
}
all_results.append(result_dict)
print(f" ✅ Completed tri-head Grad-CAM++ analysis")
except Exception as e:
print(f" ❌ Error: {e}")
import traceback
traceback.print_exc()
continue
print("\n" + "="*80)
print("✅ TRI-HEAD GRAD-CAM++ ANALYSIS COMPLETE")
print(f"📦 Generated {len(all_results)} comprehensive result dictionaries")
print("="*80)
return all_results
# -------------------------------------------
# Visualization Function
# -------------------------------------------
def display_tri_head_gradcam_grid(results):
"""
Display grid: each row = one image with 5 columns (or 4 if no stage analysis)
[Original | Disease Grad-CAM++ | Severity Grad-CAM++ | Stage Grad-CAM++ (if available) | Union]
"""
if not results:
print("No results to display")
return
num_images = len(results)
max_cols = 5
# Create figure
fig, axes = plt.subplots(num_images, max_cols, figsize=(35, 7 * num_images))
# Handle single image case
if num_images == 1:
axes = axes.reshape(1, -1)
cmap = plt.cm.jet
for i, result in enumerate(results):
has_stage = result['include_stage_analysis']
# Column 1: Original Image
axes[i, 0].imshow(result['image'])
title_text = (
f"Original Image {i+1}\n"
f"Dataset: {result['dataset_name']}\n"
f"True Label: {result['true_label']}\n"
f"Stage GT: {result['stage_label']}\n"
f"File: {result['filename']}"
)
axes[i, 0].set_title(title_text, fontsize=10, fontweight='bold', pad=10)
axes[i, 0].axis('off')
# Column 2: Disease Head Grad-CAM++ (Level 1)
axes[i, 1].imshow(result['image'])
disease_heatmap = result['level1_disease']['heatmap_raw']
im1 = axes[i, 1].imshow(
disease_heatmap,
cmap=cmap,
alpha=0.5,
vmin=0,
vmax=1
)
cbar1 = plt.colorbar(im1, ax=axes[i, 1], fraction=0.046, pad=0.04)
cbar1.set_label('Activation', rotation=270, labelpad=15)
disease_title = (
f"Level 1: Disease Head\n"
f"Predicted: {result['level1_disease']['predicted_class']}\n"
f"Confidence: {result['level1_disease']['confidence']:.4f}\n"
f"Mean Act: {result['level1_disease']['activation_stats']['mean']:.4f}"
)
axes[i, 1].set_title(disease_title, fontsize=10, fontweight='bold', pad=10)
axes[i, 1].axis('off')
# Column 3: Severity Head Grad-CAM++ (Level 2)
axes[i, 2].imshow(result['image'])
severity_heatmap = result['level2_severity']['heatmap_raw']
im2 = axes[i, 2].imshow(
severity_heatmap,
cmap=cmap,
alpha=0.5,
vmin=0,
vmax=1
)
cbar2 = plt.colorbar(im2, ax=axes[i, 2], fraction=0.046, pad=0.04)
cbar2.set_label('Activation', rotation=270, labelpad=15)
severity_title = (
f"Level 2: Severity Head\n"
f"Predicted: {result['level2_severity']['predicted_class']}\n"
f"Confidence: {result['level2_severity']['confidence']:.4f}\n"
f"Mean Act: {result['level2_severity']['activation_stats']['mean']:.4f}"
)
axes[i, 2].set_title(severity_title, fontsize=10, fontweight='bold', pad=10)
axes[i, 2].axis('off')
# Column 4: Stage Head Grad-CAM++ (Level 3) - Conditional
if has_stage:
axes[i, 3].imshow(result['image'])
stage_heatmap = result['level3_stage']['heatmap_raw']
im3 = axes[i, 3].imshow(
stage_heatmap,
cmap=cmap,
alpha=0.5,
vmin=0,
vmax=1
)
cbar3 = plt.colorbar(im3, ax=axes[i, 3], fraction=0.046, pad=0.04)
cbar3.set_label('Activation', rotation=270, labelpad=15)
stage_title = (
f"Level 3: Stage Head\n"
f"Predicted: Stage {result['level3_stage']['predicted_idx']}\n"
f"Confidence: {result['level3_stage']['confidence']:.4f}\n"
f"Mean Act: {result['level3_stage']['activation_stats']['mean']:.4f}"
)
axes[i, 3].set_title(stage_title, fontsize=10, fontweight='bold', pad=10)
axes[i, 3].axis('off')
else:
# Display placeholder text
axes[i, 3].text(
0.5, 0.5,
"Stage Analysis\nNot Applicable\n\n" +
(f"Reason: GT={result['true_label']}\n" if result['true_label'] != 'abnormal' else "") +
(f"Stage={result['stage_label']}" if result['stage_label'] == -1 else ""),
ha='center', va='center',
fontsize=12, color='gray',
transform=axes[i, 3].transAxes
)
axes[i, 3].axis('off')
# Column 5: Union Grad-CAM++
axes[i, 4].imshow(result['image'])
union_heatmap = result['union']['heatmap_raw']
im4 = axes[i, 4].imshow(
union_heatmap,
cmap=cmap,
alpha=0.5,
vmin=0,
vmax=1
)
cbar4 = plt.colorbar(im4, ax=axes[i, 4], fraction=0.046, pad=0.04)
cbar4.set_label('Activation', rotation=270, labelpad=15)
union_title = (
f"Union: Combined Grad-CAM++\n"
f"Averaged {result['union']['num_heads_averaged']} Heads\n"
f"Disease: {result['level1_disease']['predicted_class']}\n"
f"Severity: {result['level2_severity']['predicted_class']}"
)
if has_stage:
union_title += f"\nStage: {result['level3_stage']['predicted_idx']}"
union_title += f"\nMean Act: {result['union']['activation_stats']['mean']:.4f}"
axes[i, 4].set_title(union_title, fontsize=10, fontweight='bold', pad=10)
axes[i, 4].axis('off')
# Print statistics
print(f"\n📊 Image {i+1} ({result['filename']}) Statistics:")
print(f" Disease Head: {result['level1_disease']['predicted_class']} "
f"({result['level1_disease']['confidence']:.4f})")
print(f" Mean Activation: {result['level1_disease']['activation_stats']['mean']:.4f}, "
f"Max: {result['level1_disease']['activation_stats']['max']:.4f}")
print(f" Severity Head: {result['level2_severity']['predicted_class']} "
f"({result['level2_severity']['confidence']:.4f})")
print(f" Mean Activation: {result['level2_severity']['activation_stats']['mean']:.4f}, "
f"Max: {result['level2_severity']['activation_stats']['max']:.4f}")
if has_stage:
print(f" Stage Head: Stage {result['level3_stage']['predicted_idx']} "
f"({result['level3_stage']['confidence']:.4f})")
print(f" Mean Activation: {result['level3_stage']['activation_stats']['mean']:.4f}, "
f"Max: {result['level3_stage']['activation_stats']['max']:.4f}")
else:
print(f" Stage Head: Not analyzed (GT: {result['true_label']}, Stage: {result['stage_label']})")
print(f" Union Heatmap ({result['union']['num_heads_averaged']} heads):")
print(f" Mean Activation: {result['union']['activation_stats']['mean']:.4f}, "
f"Max: {result['union']['activation_stats']['max']:.4f}")
plt.suptitle(
'Hierarchical Model - Tri-Head Grad-CAM++ Analysis with Union\n'
'Level 1: Disease | Level 2: Severity | Level 3: Stage (Conditional) | Union: Combined Analysis\n'
'Red = High Activation | Blue = Low Activation',
fontsize=16,
fontweight='bold',
y=0.998
)
plt.tight_layout()
# Save
grid_save_path = os.path.join(GRADCAM_OUTPUT_DIR, 'tri_head_union_gradcam_plus_plus_analysis.png')
plt.savefig(grid_save_path, dpi=150, bbox_inches='tight')
print(f"\n✅ Grid saved to: {grid_save_path}")
plt.show()
# -------------------------------------------
# Execute Analysis
# -------------------------------------------
print("\n" + "="*80)
print("CHECKING FOR COLLECTED IMAGES")
print("="*80)
try:
if 'collected_images' in locals() or 'collected_images' in globals():
print(f"✅ Found collected_images with {len(collected_images)} images\n")
# Run tri-head Grad-CAM++ analysis
gradcam_results = run_tri_head_gradcam_plus_plus_analysis(
model,
device,
collected_images
)
# Display results
if gradcam_results:
display_tri_head_gradcam_grid(gradcam_results)
# Save summary
results_summary = []
for r in gradcam_results:
summary_row = {
'filename': r['filename'],
'dataset': r['dataset_name'],
'true_label': r['true_label'],
'stage_gt': r['stage_label'],
'disease_predicted': r['level1_disease']['predicted_class'],
'disease_confidence': r['level1_disease']['confidence'],
'disease_mean_activation': r['level1_disease']['activation_stats']['mean'],
'severity_predicted': r['level2_severity']['predicted_class'],
'severity_confidence': r['level2_severity']['confidence'],
'severity_mean_activation': r['level2_severity']['activation_stats']['mean'],
}
if r['include_stage_analysis']:
summary_row.update({
'stage_predicted': r['level3_stage']['predicted_idx'],
'stage_confidence': r['level3_stage']['confidence'],
'stage_mean_activation': r['level3_stage']['activation_stats']['mean'],
})
else:
summary_row.update({
'stage_predicted': 'N/A',
'stage_confidence': 'N/A',
'stage_mean_activation': 'N/A',
})
summary_row['union_heads_averaged'] = r['union']['num_heads_averaged']
summary_row['union_mean_activation'] = r['union']['activation_stats']['mean']
results_summary.append(summary_row)
summary_df = pd.DataFrame(results_summary)
summary_path = os.path.join(GRADCAM_OUTPUT_DIR, 'tri_head_union_gradcam_plus_plus_summary.csv')
summary_df.to_csv(summary_path, index=False)
print(f"\n✅ Summary saved to: {summary_path}")
print("\n" + "="*80)
print("TRI-HEAD GRAD-CAM++ ANALYSIS SUMMARY")
print("="*80)
print(summary_df.to_string(index=False))
print("="*80)
# Print structure of results for reference
print("\n" + "="*80)
print("📦 RESULTS STRUCTURE")
print("="*80)
print("Each result dictionary contains:")
print(" - filename, full_path, dataset_name, true_label, stage_label")
print(" - include_stage_analysis: boolean flag")
print(" - image: original image array")
print(" - level1_disease: {")
print(" predicted_class, predicted_idx, confidence, all_probabilities")
print(" heatmap_raw, heatmap_overlay, activation_stats")
print(" }")
print(" - level2_severity: {")
print(" predicted_class, predicted_idx, confidence, all_probabilities")
print(" heatmap_raw, heatmap_overlay, activation_stats")
print(" }")
print(" - level3_stage: {")
print(" predicted_idx, confidence, all_probabilities")
print(" heatmap_raw (None if not analyzed), heatmap_overlay (None if not analyzed)")
print(" activation_stats (None if not analyzed)")
print(" }")
print(" - union: {")
print(" heatmap_raw (average of 2 or 3 heads)")
print(" heatmap_overlay, num_heads_averaged, activation_stats")
print(" }")
print("="*80)
print(f"\n✅ gradcam_results variable contains {len(gradcam_results)} dictionaries")
print(" Use gradcam_results in the next cell for further analysis!")
else:
print("\n❌ No results generated")
else:
print("❌ collected_images not found!")
except Exception as e:
print(f"❌ Error: {e}")
import traceback
traceback.print_exc()
import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import cv2
from scipy.ndimage import zoom, gaussian_filter
import os
# -------------------------------------------
# Configuration
# -------------------------------------------
OUTPUT_DIR="/kaggle/working"
ATTENTION_OUTPUT_DIR = os.path.join(OUTPUT_DIR, 'attention_results')
os.makedirs(ATTENTION_OUTPUT_DIR, exist_ok=True)
print("="*80)
print("ATTENTION VISUALIZATION CONFIGURATION")
print("="*80)
print(f"Output directory: {ATTENTION_OUTPUT_DIR}")
print("="*80)
# -------------------------------------------
# Attention Extraction Wrapper
# -------------------------------------------
class AttentionExtractor(nn.Module):
"""
Wrapper to extract attention weights from the hierarchical model
"""
def __init__(self, phase3_model):
super().__init__()
self.phase3_model = phase3_model
self.attention_weights = None
self.tile_features = None
def forward(self, tiles):
"""
Extract attention weights and tile features from the model
"""
# Get model outputs including attention weights
disease_logits, severity_logits, stage_logits, attention_weights = self.phase3_model(tiles)
# Store attention weights for visualization
self.attention_weights = attention_weights
return disease_logits, severity_logits, stage_logits, attention_weights
# Create attention extractor
attention_extractor = AttentionExtractor(model).to(device)
attention_extractor.eval()
print("\n✅ Attention extractor created successfully\n")
# -------------------------------------------
# Helper Functions
# -------------------------------------------
def extract_attention_map(model, preprocessed_image, device):
"""
Extract attention weights from the model for a single image
Since each image is a single tile, we get attention for that single representation
Args:
model: AttentionExtractor model
preprocessed_image: Preprocessed tensor (single image, not tiled)
device: torch device
Returns:
attention_weights: numpy array of attention weight (single value for single tile)
disease_logits, severity_logits, stage_logits: model outputs
"""
model.eval()
try:
# Each preprocessed image is already a single tensor of shape (C, H, W)
# We need to add batch dimension and wrap in list
if preprocessed_image.dim() == 3:
# Single image: (C, H, W) -> (1, C, H, W)
image_batch = preprocessed_image.unsqueeze(0)
else:
# Already has batch dimension
image_batch = preprocessed_image
# Wrap in list as model expects list of tile batches
# Since we have single image as single tile, this is [1 tile batch]
tiles_list = [image_batch.to(device)]
with torch.no_grad():
disease_logits, severity_logits, stage_logits, attention_weights = model(tiles_list)
# Convert attention weights to numpy
# For single tile, this will be shape (1,) or (1, 1)
attention_np = attention_weights.squeeze().cpu().numpy()
# Ensure it's at least 1D
if attention_np.ndim == 0:
attention_np = np.array([attention_np.item()])
print(f" Extracted attention weights: shape={attention_np.shape}, value={attention_np}")
return attention_np, disease_logits, severity_logits, stage_logits
except Exception as e:
print(f" Error extracting attention: {e}")
import traceback
traceback.print_exc()
return None, None, None, None
def create_uniform_attention_heatmap(attention_weight, image_shape):
"""
Create a uniform attention heatmap for a single tile (entire image)
Since the whole image is one tile, the attention is uniform across it
Args:
attention_weight: single attention weight value
image_shape: tuple (height, width) of image
Returns:
heatmap: 2D array with uniform attention value
"""
# Since we have single tile = whole image, create uniform heatmap
# with the attention weight value
heatmap = np.full(image_shape, attention_weight, dtype=np.float32)
# Normalize to [0, 1] for visualization
if heatmap.max() > 0:
heatmap = heatmap / heatmap.max()
return heatmap
def extract_patch_level_attention(model_backbone, preprocessed_image, device, patch_size=16):
"""
Extract patch-level attention from ViT backbone
ViT processes image as patches, we can visualize their importance
Args:
model_backbone: ViT backbone model
preprocessed_image: Preprocessed tensor
device: torch device
patch_size: ViT patch size (default 16 for most ViTs)
Returns:
patch_attention_map: 2D heatmap showing patch-level importance
"""
try:
if preprocessed_image.dim() == 3:
image_batch = preprocessed_image.unsqueeze(0).to(device)
else:
image_batch = preprocessed_image.to(device)
with torch.no_grad():
# Get ViT outputs - last_hidden_state contains all patch embeddings
outputs = model_backbone.vit(pixel_values=image_batch)
# Shape: (batch, num_patches + 1, embed_dim)
# First token is CLS token, rest are patch tokens
hidden_states = outputs.last_hidden_state
# Get patch tokens (exclude CLS token at index 0)
patch_tokens = hidden_states[:, 1:, :] # (1, num_patches, embed_dim)
# Compute importance as L2 norm of each patch embedding
patch_importance = torch.norm(patch_tokens, p=2, dim=2).squeeze().cpu().numpy()
# Calculate grid dimensions
# For 224x224 image with patch_size=16: 14x14 = 196 patches
num_patches = len(patch_importance)
grid_size = int(np.sqrt(num_patches))
# Reshape to 2D grid
attention_grid = patch_importance.reshape(grid_size, grid_size)
print(f" Extracted patch-level attention: {grid_size}x{grid_size} patches")
return attention_grid
except Exception as e:
print(f" Error extracting patch attention: {e}")
import traceback
traceback.print_exc()
return None
def create_patch_attention_heatmap(patch_attention_grid, target_shape):
"""
Upsample patch-level attention to image dimensions
Args:
patch_attention_grid: 2D grid of patch attention values
target_shape: tuple (height, width) for output
Returns:
heatmap: upsampled attention heatmap
"""
# Calculate zoom factors
zoom_factors = (target_shape[0] / patch_attention_grid.shape[0],
target_shape[1] / patch_attention_grid.shape[1])
# Upsample using bilinear interpolation
heatmap = zoom(patch_attention_grid, zoom_factors, order=1)
# Normalize to [0, 1]
heatmap = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min() + 1e-8)
# Apply smoothing for better visualization
heatmap = gaussian_filter(heatmap, sigma=5)
return heatmap
def load_image_for_attention(img_path, target_size=768):
"""Load and resize image for attention visualization"""
try:
if img_path.lower().endswith(('.svs', '.ndpi')):
slide = openslide.OpenSlide(img_path)
img_pil = slide.get_thumbnail((target_size, target_size))
slide.close()
elif img_path.lower().endswith('.tif'):
try:
slide = openslide.OpenSlide(img_path)
img_pil = slide.get_thumbnail((target_size, target_size))
slide.close()
except:
img_pil = Image.open(img_path).convert('RGB')
img_pil = img_pil.resize((target_size, target_size), Image.BILINEAR)
else:
img_pil = Image.open(img_path).convert('RGB')
img_pil = img_pil.resize((target_size, target_size), Image.BILINEAR)
img_array = np.array(img_pil)
return img_pil, img_array
except Exception as e:
print(f"Error loading image {img_path}: {e}")
return None, None
# -------------------------------------------
# Main Attention Extraction Function
# -------------------------------------------
def run_attention_analysis(attention_model, device, collected_images, processed_images,
main_class_mapping, stage_class_mapping):
"""
Extract and visualize attention weights from the hierarchical model
Uses patch-level attention from ViT backbone since images are single tiles
Args:
attention_model: AttentionExtractor model
device: torch device
collected_images: list of image paths
processed_images: list of preprocessed tensors (single images)
main_class_mapping: dictionary mapping class indices to names
stage_class_mapping: dictionary mapping stage indices to names
Returns:
list of results dictionaries
"""
print("\n" + "="*80)
print("🎯 STARTING ATTENTION WEIGHT EXTRACTION AND VISUALIZATION")
print("="*80)
print("ℹ️ Note: Each image is treated as a single tile")
print("ℹ️ Using patch-level attention from ViT backbone for visualization")
if not collected_images or not processed_images:
print("❌ No images or preprocessed data provided")
return []
print(f"✅ Processing {len(collected_images)} images\n")
all_results = []
for idx, (img_path, preprocessed_image) in enumerate(zip(collected_images, processed_images)):
print(f"\n[{idx+1}/{len(collected_images)}] Processing: {os.path.basename(img_path)}")
print(f" Path: {img_path}")
print(f" Image shape: {preprocessed_image.shape}")
try:
# Load original image for visualization
img_pil, img_array = load_image_for_attention(img_path, target_size=768)
if img_array is None:
print(f" ❌ Failed to load image")
continue
# Extract MIL-level attention weights (single value for single tile)
attention_weights, disease_logits, severity_logits, stage_logits = extract_attention_map(
attention_model,
preprocessed_image,
device
)
if attention_weights is None:
print(f" ❌ Failed to extract attention")
continue
# Extract patch-level attention from ViT backbone
patch_attention = extract_patch_level_attention(
attention_model.phase3_model.backbone,
preprocessed_image,
device
)
if patch_attention is not None:
# Create heatmap from patch attention
attention_heatmap = create_patch_attention_heatmap(
patch_attention,
img_array.shape[:2]
)
print(f" ✅ Created patch-level attention heatmap")
else:
# Fallback: uniform heatmap with MIL attention weight
attention_heatmap = create_uniform_attention_heatmap(
attention_weights[0],
img_array.shape[:2]
)
print(f" ℹ️ Using uniform attention heatmap")
# Get predictions
with torch.no_grad():
disease_probs = F.softmax(disease_logits, dim=1)
disease_pred_idx = torch.argmax(disease_probs, dim=1).item()
disease_confidence = disease_probs[0, disease_pred_idx].item()
predicted_class_name = main_class_mapping.get(
disease_pred_idx,
f"Unknown_Class_{disease_pred_idx}"
)
# Get severity prediction
if "_normal" in predicted_class_name:
predicted_disease = predicted_class_name.replace("_normal", "")
severity_label = "Normal"
elif "_abnormal" in predicted_class_name:
predicted_disease = predicted_class_name.replace("_abnormal", "")
severity_label = "Abnormal"
else:
predicted_disease = predicted_class_name
severity_label = "Unknown"
print(f" 📊 Prediction: {predicted_class_name}")
print(f" Confidence: {disease_confidence:.4f}")
print(f" MIL Attention Weight: {attention_weights[0]:.4f}")
# Extract dataset and true label from path
path_parts = img_path.split(os.sep)
dataset_name = "Unknown"
true_label = "unknown"
for part in path_parts:
if "processed" in part or "cancer" in part.lower():
dataset_name = part
if part in ["normal", "abnormal"]:
true_label = part
break
# Store results
result = {
'image': img_array,
'attention_heatmap': attention_heatmap,
'mil_attention_weight': attention_weights[0],
'true_label': true_label,
'dataset_name': dataset_name,
'predicted_class': predicted_class_name,
'predicted_disease': predicted_disease,
'severity': severity_label,
'class_idx': disease_pred_idx,
'confidence': disease_confidence,
'filename': os.path.basename(img_path),
'full_path': img_path
}
all_results.append(result)
# Save individual attention heatmap (raw, for OpenCV processing)
heatmap_filename = f"attention_heatmap_{idx+1}_{os.path.splitext(os.path.basename(img_path))[0]}.npy"
heatmap_path = os.path.join(ATTENTION_OUTPUT_DIR, heatmap_filename)
np.save(heatmap_path, attention_heatmap)
print(f" ✅ Attention heatmap saved to: {heatmap_filename}")
print(f" ✅ Completed analysis")
except Exception as e:
print(f" ❌ Error: {e}")
import traceback
traceback.print_exc()
continue
print("\n" + "="*80)
print("✅ ATTENTION EXTRACTION COMPLETE")
print(f"📁 Results saved to: {ATTENTION_OUTPUT_DIR}")
print("="*80)
return all_results
# -------------------------------------------
# Visualization Function
# -------------------------------------------
def display_attention_grid(results):
"""
Display grid with original images and attention heatmap overlays
"""
if not results:
print("No results to display")
return
num_images = len(results)
# Create figure: 3 columns (original, heatmap, overlay)
fig, axes = plt.subplots(num_images, 3, figsize=(18, 6 * num_images))
# Handle single image case
if num_images == 1:
axes = axes.reshape(1, -1)
# Use 'jet' colormap for attention (blue to red)
cmap = plt.cm.jet
for i, result in enumerate(results):
# Column 1: Original Image
axes[i, 0].imshow(result['image'])
title_text = (
f"Original Image {i+1}\n"
f"Dataset: {result['dataset_name']}\n"
f"True Label: {result['true_label']}\n"
f"File: {result['filename'][:30]}..."
)
axes[i, 0].set_title(title_text, fontsize=10, fontweight='bold', pad=10)
axes[i, 0].axis('off')
# Column 2: Attention Heatmap
im = axes[i, 1].imshow(result['attention_heatmap'], cmap=cmap)
cbar = plt.colorbar(im, ax=axes[i, 1], fraction=0.046, pad=0.04)
cbar.set_label('Attention Weight', rotation=270, labelpad=15)
heatmap_title = (
f"Attention Heatmap {i+1}\n"
f"Patch-Level Importance\n"
f"MIL Weight: {result['mil_attention_weight']:.4f}"
)
axes[i, 1].set_title(heatmap_title, fontsize=10, fontweight='bold', pad=10)
axes[i, 1].axis('off')
# Column 3: Overlay
axes[i, 2].imshow(result['image'])
axes[i, 2].imshow(result['attention_heatmap'], cmap=cmap, alpha=0.5)
overlay_title = (
f"Overlay {i+1}\n"
f"Predicted: {result['predicted_class']}\n"
f"Confidence: {result['confidence']:.4f}"
)
axes[i, 2].set_title(overlay_title, fontsize=10, fontweight='bold', pad=10)
axes[i, 2].axis('off')
# Print statistics
high_attention = np.sum(result['attention_heatmap'] > 0.7) / result['attention_heatmap'].size * 100
medium_attention = np.sum((result['attention_heatmap'] > 0.4) &
(result['attention_heatmap'] <= 0.7)) / result['attention_heatmap'].size * 100
low_attention = np.sum(result['attention_heatmap'] <= 0.4) / result['attention_heatmap'].size * 100
print(f"\n📊 Image {i+1} ({result['filename']}) Attention Statistics:")
print(f" Predicted: {result['predicted_class']}")
print(f" Confidence: {result['confidence']:.4f}")
print(f" MIL Attention Weight: {result['mil_attention_weight']:.4f}")
print(f" High attention regions (>0.7): {high_attention:.1f}%")
print(f" Medium attention regions (0.4-0.7): {medium_attention:.1f}%")
print(f" Low attention regions (<0.4): {low_attention:.1f}%")
plt.suptitle(
'Hierarchical Model - Patch-Level Attention Visualization\n'
'Warmer colors (red/yellow) indicate higher attention | Cooler colors (blue) indicate lower attention',
fontsize=16,
fontweight='bold',
y=0.998
)
plt.tight_layout()
# Save grid
grid_save_path = os.path.join(ATTENTION_OUTPUT_DIR, 'attention_visualization_grid.png')
plt.savefig(grid_save_path, dpi=150, bbox_inches='tight')
print(f"\n✅ Grid visualization saved to: {grid_save_path}")
plt.show()
# -------------------------------------------
# Execute Attention Analysis
# -------------------------------------------
print("\n" + "="*80)
print("CHECKING FOR COLLECTED AND PROCESSED IMAGES")
print("="*80)
try:
# Check if required variables exist
if 'collected_images' in locals() or 'collected_images' in globals():
if 'processed_images' in locals() or 'processed_images' in globals():
print(f"✅ Found collected_images: {len(collected_images)} images")
print(f"✅ Found processed_images: {len(processed_images)} tensors\n")
# Run attention analysis
attention_results = run_attention_analysis(
attention_extractor,
device,
collected_images,
processed_images,
DISEASE_CLASS_MAPPING,
STAGE_CLASS_MAPPING
)
# Display results
if attention_results:
display_attention_grid(attention_results)
# Save results summary
results_summary = []
for r in attention_results:
high_attn = np.sum(r['attention_heatmap'] > 0.7) / r['attention_heatmap'].size * 100
results_summary.append({
'filename': r['filename'],
'dataset': r['dataset_name'],
'true_label': r['true_label'],
'predicted_class': r['predicted_class'],
'confidence': r['confidence'],
'mil_attention_weight': r['mil_attention_weight'],
'high_attention_area_%': high_attn,
'max_attention': r['attention_heatmap'].max(),
'mean_attention': r['attention_heatmap'].mean()
})
summary_df = pd.DataFrame(results_summary)
summary_path = os.path.join(ATTENTION_OUTPUT_DIR, 'attention_summary.csv')
summary_df.to_csv(summary_path, index=False)
print(f"\n✅ Summary saved to: {summary_path}")
print("\n" + "="*80)
print("ATTENTION ANALYSIS SUMMARY")
print("="*80)
print(summary_df.to_string(index=False))
print("="*80)
# Save attention results for next cell (OpenCV feature extraction)
print("\n" + "="*80)
print("ATTENTION RESULTS READY FOR OPENCV PROCESSING")
print("="*80)
print(f"✅ Variable 'attention_results' contains {len(attention_results)} results")
print("✅ Each result includes:")
print(" - Original image")
print(" - Patch-level attention heatmap (smoothed)")
print(" - MIL attention weight")
print(" - Predictions and metadata")
print("\n💡 Use 'attention_results' in the next cell for OpenCV feature extraction")
print("="*80)
else:
print("\n❌ No results generated")
else:
print("❌ processed_images not found!")
print("Please run the preprocessing cell first")
else:
print("❌ collected_images not found!")
print("Please run the image collection cell first")
except Exception as e:
print(f"❌ Error: {e}")
import traceback
traceback.print_exc()
import numpy as np
import cv2
from sklearn.cluster import DBSCAN
import matplotlib.pyplot as plt
import os
from scipy.ndimage import maximum_filter
from skimage.feature import graycomatrix, graycoprops
from openai import OpenAI
import time
from scipy.stats import spearmanr
# ================================================================
# HEATMAP FEATURE EXTRACTOR CLASS
# Adapted for Hierarchical Model Pipeline
# ================================================================
class HeatmapFeatureExtractor:
def __init__(self, attention_result):
"""
attention_result is one entry from attention_results
"""
self.heatmap = attention_result['attention_heatmap']
self.original_image = attention_result['image']
self.prediction_info = {
'predicted_class': attention_result['predicted_class'],
'predicted_disease': attention_result['predicted_disease'],
'severity': attention_result['severity'],
'confidence': attention_result['confidence'],
'class_idx': attention_result['class_idx'],
'mil_attention': attention_result['mil_attention_weight']
}
self.true_label = attention_result['true_label']
self.dataset_name = attention_result['dataset_name']
self.filename = attention_result['filename']
# ---------------------------------------------------------
# METHOD 1: Brightest Region Analysis
# ---------------------------------------------------------
def get_brightest_region(self):
"""
IMPROVED: Comprehensive analysis of high-attention regions
"""
heatmap = self.heatmap.astype(float)
H, W = heatmap.shape
heatmap_norm = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min() + 1e-8)
# 1. PRIMARY HOTSPOT
brightest_idx = np.unravel_index(np.argmax(heatmap_norm), heatmap_norm.shape)
y_bright, x_bright = brightest_idx
intensity_bright = heatmap_norm[y_bright, x_bright]
position_bright = self._get_anatomical_position(y_bright, x_bright, H, W)
primary_hotspot = {
"pixel": (int(y_bright), int(x_bright)),
"position": position_bright,
"intensity": float(intensity_bright)
}
# 2. SECONDARY HOTSPOTS
secondary_hotspots = self._find_secondary_hotspots(heatmap_norm, H, W, threshold=0.6)
# 3. ATTENTION PATTERN
attention_pattern = self._determine_attention_pattern(heatmap_norm, H, W)
# 4. SPATIAL COVERAGE
spatial_coverage = self._calculate_spatial_coverage(heatmap_norm, H, W)
# 5. HOTSPOT COUNT
hotspot_count = 1 + len(secondary_hotspots)
return {
"primary_hotspot": primary_hotspot,
"secondary_hotspots": secondary_hotspots,
"attention_pattern": attention_pattern,
"spatial_coverage": spatial_coverage,
"hotspot_count": hotspot_count
}
def _get_anatomical_position(self, y, x, H, W):
"""Convert pixel coordinates to descriptive position"""
y_rel = y / H
x_rel = x / W
center_y, center_x = H / 2, W / 2
dist_from_center = np.sqrt((y - center_y)**2 + (x - center_x)**2)
max_dist = np.sqrt((H/2)**2 + (W/2)**2)
dist_ratio = dist_from_center / max_dist
center_threshold_inner = 0.35
center_threshold_outer = 0.65
periphery_threshold = 0.75
if y_rel < center_threshold_inner:
vert = "upper"
elif y_rel > center_threshold_outer:
vert = "lower"
else:
vert = "mid"
if x_rel < center_threshold_inner:
horiz = "left"
elif x_rel > center_threshold_outer:
horiz = "right"
else:
horiz = "center"
if horiz == "center" and vert == "mid":
position = "center"
elif horiz == "center":
position = f"{vert}-center"
elif vert == "mid":
position = f"{horiz}-center"
else:
position = f"{vert}-{horiz}"
if dist_ratio > periphery_threshold:
position = f"{position} (periphery)"
elif dist_ratio < 0.3:
position = f"{position} (core)"
return position
def _find_secondary_hotspots(self, heatmap_norm, H, W, threshold=0.6, min_distance=20):
"""Find additional significant attention regions"""
secondary = []
neighborhood_size = max(10, min(H, W) // 20)
local_max = maximum_filter(heatmap_norm, size=neighborhood_size)
peaks = (heatmap_norm == local_max) & (heatmap_norm > threshold * heatmap_norm.max())
peak_coords = np.argwhere(peaks)
peak_intensities = heatmap_norm[peaks]
sorted_indices = np.argsort(peak_intensities)[::-1]
primary_y, primary_x = np.unravel_index(np.argmax(heatmap_norm), heatmap_norm.shape)
for idx in sorted_indices[:5]:
y, x = peak_coords[idx]
if np.sqrt((y - primary_y)**2 + (x - primary_x)**2) < min_distance:
continue
too_close = False
for existing in secondary:
ey, ex = existing["pixel"]
if np.sqrt((y - ey)**2 + (x - ex)**2) < min_distance:
too_close = True
break
if too_close:
continue
position = self._get_anatomical_position(y, x, H, W)
intensity = float(heatmap_norm[y, x])
secondary.append({
"pixel": (int(y), int(x)),
"position": position,
"intensity": intensity
})
return secondary
def _determine_attention_pattern(self, heatmap_norm, H, W):
"""Determine overall attention distribution pattern"""
center_y, center_x = H // 2, W // 2
Y, X = np.ogrid[:H, :W]
dist_from_center = np.sqrt((Y - center_y)**2 + (X - center_x)**2)
max_dist = np.sqrt((H/2)**2 + (W/2)**2)
core_mask = dist_from_center < (max_dist * 0.3)
mid_mask = (dist_from_center >= max_dist * 0.3) & (dist_from_center < max_dist * 0.7)
periphery_mask = dist_from_center >= (max_dist * 0.7)
core_attention = np.mean(heatmap_norm[core_mask])
mid_attention = np.mean(heatmap_norm[mid_mask])
periphery_attention = np.mean(heatmap_norm[periphery_mask])
high_attention_pixels = np.sum(heatmap_norm > 0.7) / heatmap_norm.size
if core_attention > 0.7 and core_attention > mid_attention * 1.5:
return "centralized (focused on center)"
elif periphery_attention > 0.7 and periphery_attention > core_attention * 1.5:
return "peripheral (focused on edges)"
elif mid_attention > core_attention and mid_attention > periphery_attention:
return "ring-like (donut pattern)"
elif high_attention_pixels > 0.5:
return "diffuse (spread across image)"
elif high_attention_pixels > 0.1 and high_attention_pixels < 0.3:
return "focal (single concentrated region)"
else:
return "scattered (multiple regions)"
def _calculate_spatial_coverage(self, heatmap_norm, H, W):
"""Calculate percentage of attention in each spatial region"""
center_y, center_x = H // 2, W // 2
Y, X = np.ogrid[:H, :W]
dist_from_center = np.sqrt((Y - center_y)**2 + (X - center_x)**2)
max_dist = np.sqrt((H/2)**2 + (W/2)**2)
core_mask = dist_from_center < (max_dist * 0.3)
mid_mask = (dist_from_center >= max_dist * 0.3) & (dist_from_center < max_dist * 0.7)
periphery_mask = dist_from_center >= (max_dist * 0.7)
total_attention = np.sum(heatmap_norm)
core_sum = np.sum(heatmap_norm[core_mask])
mid_sum = np.sum(heatmap_norm[mid_mask])
periphery_sum = np.sum(heatmap_norm[periphery_mask])
return {
"center_attention": float(core_sum / total_attention * 100) if total_attention > 0 else 0,
"mid_region_attention": float(mid_sum / total_attention * 100) if total_attention > 0 else 0,
"periphery_attention": float(periphery_sum / total_attention * 100) if total_attention > 0 else 0
}
# ---------------------------------------------------------
# METHOD 2: Scatter Analysis
# ---------------------------------------------------------
def get_activation_scatter(self, threshold_ratio=0.6):
"""Determine if heatmap is focused or scattered"""
heatmap = self.heatmap.astype(float)
H, W = heatmap.shape
heatmap_norm = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min() + 1e-6)
thresh = threshold_ratio * heatmap_norm.max()
high_pixels = np.argwhere(heatmap_norm >= thresh)
if len(high_pixels) == 0:
return {
"scatter_level": "low",
"num_clusters": 0,
"clusters_sizes": []
}
clustering = DBSCAN(eps=8, min_samples=20).fit(high_pixels)
labels = clustering.labels_
unique_labels = [lb for lb in np.unique(labels) if lb != -1]
cluster_sizes = []
for lb in unique_labels:
cluster_sizes.append(int(np.sum(labels == lb)))
num_clusters = len(unique_labels)
if num_clusters == 1:
scatter = "low"
elif 2 <= num_clusters <= 3:
scatter = "medium"
else:
scatter = "high"
return {
"scatter_level": scatter,
"num_clusters": num_clusters,
"clusters_sizes": cluster_sizes
}
# ---------------------------------------------------------
# METHOD 3: Dominant Color Analysis
# ---------------------------------------------------------
def get_dominant_focus_color(self, threshold_ratio=0.6, k_clusters=5):
"""Detect dominant color in attention-focused regions"""
heatmap = self.heatmap.astype(float)
orig = self.original_image.copy()
H, W = heatmap.shape
heatmap_norm = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min() + 1e-6)
mask = (heatmap_norm >= threshold_ratio).astype(np.uint8) * 255
if np.sum(mask) == 0:
return {
"dominant_color_rgb": None,
"dominant_color_hsv": None,
"dominant_color_name": "none",
"color_confidence": 0.0
}
kernel = np.ones((3, 3), np.uint8)
mask_clean = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel, iterations=2)
contours, _ = cv2.findContours(mask_clean, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if len(contours) == 0:
return {
"dominant_color_rgb": None,
"dominant_color_hsv": None,
"dominant_color_name": "none",
"color_confidence": 0.0
}
activation_mask = np.zeros_like(mask_clean)
cv2.drawContours(activation_mask, contours, -1, 255, -1)
focus_pixels = orig[activation_mask == 255]
if len(focus_pixels) < 10:
return {
"dominant_color_rgb": None,
"dominant_color_hsv": None,
"dominant_color_name": "none",
"color_confidence": 0.0
}
Z = np.float32(focus_pixels)
criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 20, 1.0)
K = k_clusters
_, labels, centers = cv2.kmeans(
Z, K, None, criteria, 10, cv2.KMEANS_RANDOM_CENTERS
)
counts = np.bincount(labels.flatten())
sorted_indices = np.argsort(counts)[::-1]
dominant_index = None
dominant_color = None
for idx in sorted_indices:
candidate_color = centers[idx].astype(int)
hsv_color = cv2.cvtColor(
np.uint8([[candidate_color]]),
cv2.COLOR_RGB2HSV
)[0][0]
if hsv_color[1] > 30:
dominant_index = idx
dominant_color = candidate_color
break
if dominant_index is None:
dominant_index = sorted_indices[0]
dominant_color = centers[dominant_index].astype(int)
hsv_color = cv2.cvtColor(
np.uint8([[dominant_color]]),
cv2.COLOR_RGB2HSV
)[0][0]
color_confidence = counts[dominant_index] / len(labels) * 100
dominant_name = self._map_color_to_name(
dominant_color.tolist(),
hsv_color.tolist()
)
return {
"dominant_color_rgb": dominant_color.tolist(),
"dominant_color_hsv": hsv_color.tolist(),
"dominant_color_name": dominant_name,
"color_confidence": float(color_confidence)
}
def _map_color_to_name(self, rgb, hsv=None):
"""Enhanced color naming using HSV color space"""
r, g, b = rgb
if hsv is None:
hsv_array = cv2.cvtColor(np.uint8([[rgb]]), cv2.COLOR_RGB2HSV)[0][0]
h, s, v = hsv_array.tolist()
else:
h, s, v = hsv
if s < 30:
if v > 200:
return "white / very light"
elif v > 150:
return "light gray / pale"
elif v > 80:
return "gray"
else:
return "dark gray / black"
if v < 60:
return "very dark / black"
if 130 <= h <= 160:
if s > 100:
return "purple / violet"
else:
return "light purple / lavender"
if 160 <= h <= 180 or h <= 10:
if v > 180 and s < 100:
return "pink / light red"
elif s > 150:
return "magenta / bright pink"
else:
return "pink / rose"
if h <= 10:
if v < 150:
return "dark red / maroon"
else:
return "red / crimson"
if 10 <= h < 25:
if v < 130:
return "brown / dark tan"
else:
return "orange / tan"
if 25 <= h < 40:
if s < 80:
return "beige / cream"
else:
return "yellow / golden"
if 40 <= h < 80:
if v > 180:
return "light green / pale green"
elif v > 120:
return "green"
else:
return "dark green"
if 80 <= h < 100:
return "cyan / turquoise"
if 100 <= h < 130:
if s > 150:
return "blue / deep blue"
elif v > 180:
return "light blue / sky blue"
else:
return "blue"
max_channel = max(r, g, b)
if max_channel == r:
return "reddish tones"
elif max_channel == g:
return "greenish tones"
elif max_channel == b:
return "bluish tones"
else:
return "mixed color region"
# ---------------------------------------------------------
# METHOD 4: Texture Analysis
# ---------------------------------------------------------
def get_texture_analysis(self, threshold_ratio=0.6):
"""
Analyze texture patterns in high-attention regions using GLCM
Returns generic image-based descriptions without medical assumptions
"""
heatmap = self.heatmap.astype(float)
orig = self.original_image.copy()
H, W = heatmap.shape
heatmap_norm = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min() + 1e-6)
mask = (heatmap_norm >= threshold_ratio).astype(np.uint8) * 255
if np.sum(mask) < 100:
return {
"texture_classification": "insufficient data",
"texture_description": "Not enough attention data to analyze texture",
"texture_scores": {
"uniformity": 0,
"organization": 0,
"complexity": 0,
"smoothness": 0
},
"glcm_features": {
"contrast": 0.0,
"correlation": 0.0,
"energy": 0.0,
"homogeneity": 0.0
}
}
# Clean mask
kernel = np.ones((3, 3), np.uint8)
mask_clean = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel, iterations=2)
# Extract focused region from original image
contours, _ = cv2.findContours(mask_clean, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if len(contours) == 0:
return {
"texture_classification": "no region detected",
"texture_description": "No focused region detected",
"texture_scores": {
"uniformity": 0,
"organization": 0,
"complexity": 0,
"smoothness": 0
},
"glcm_features": {
"contrast": 0.0,
"correlation": 0.0,
"energy": 0.0,
"homogeneity": 0.0
}
}
# Create activation mask
activation_mask = np.zeros_like(mask_clean)
cv2.drawContours(activation_mask, contours, -1, 255, -1)
# Get bounding box of the region
y_coords, x_coords = np.where(activation_mask == 255)
if len(y_coords) == 0:
return {
"texture_classification": "invalid region",
"texture_description": "Invalid region for texture analysis",
"texture_scores": {
"uniformity": 0,
"organization": 0,
"complexity": 0,
"smoothness": 0
},
"glcm_features": {
"contrast": 0.0,
"correlation": 0.0,
"energy": 0.0,
"homogeneity": 0.0
}
}
y_min, y_max = y_coords.min(), y_coords.max()
x_min, x_max = x_coords.min(), x_coords.max()
# Extract region
region_rgb = orig[y_min:y_max+1, x_min:x_max+1]
region_mask = activation_mask[y_min:y_max+1, x_min:x_max+1]
# Convert to grayscale for texture analysis
region_gray = cv2.cvtColor(region_rgb, cv2.COLOR_RGB2GRAY)
# Apply mask to focus only on high-attention pixels
region_gray_masked = region_gray.copy()
region_gray_masked[region_mask == 0] = 0
# Quantize to reduce GLCM computation (64 levels)
region_quantized = (region_gray_masked / 4).astype(np.uint8)
# Compute GLCM
# distances: [1] means immediate neighbors
# angles: [0, π/4, π/2, 3π/4] for rotation invariance
distances = [1]
angles = [0, np.pi/4, np.pi/2, 3*np.pi/4]
try:
glcm = graycomatrix(
region_quantized,
distances=distances,
angles=angles,
levels=64,
symmetric=True,
normed=True
)
# Extract features (averaged across all angles)
contrast = float(graycoprops(glcm, 'contrast')[0].mean())
correlation = float(graycoprops(glcm, 'correlation')[0].mean())
energy = float(graycoprops(glcm, 'energy')[0].mean())
homogeneity = float(graycoprops(glcm, 'homogeneity')[0].mean())
except Exception as e:
print(f" Warning: GLCM computation failed: {e}")
return {
"texture_classification": "computation error",
"texture_description": "Error computing texture features",
"texture_scores": {
"uniformity": 0,
"organization": 0,
"complexity": 0,
"smoothness": 0
},
"glcm_features": {
"contrast": 0.0,
"correlation": 0.0,
"energy": 0.0,
"homogeneity": 0.0
}
}
# Convert to 0-100 scores
uniformity_score = int(energy * 100)
organization_score = int(max(0, min(100, (correlation + 1) * 50))) # Scale -1,1 to 0,100
complexity_score = int((1 - energy) * 100)
smoothness_score = int(homogeneity * 100)
# Classify texture based on GLCM features
classification, description = self._classify_texture(
contrast, correlation, energy, homogeneity
)
return {
"texture_classification": classification,
"texture_description": description,
"texture_scores": {
"uniformity": uniformity_score,
"organization": organization_score,
"complexity": complexity_score,
"smoothness": smoothness_score
},
"glcm_features": {
"contrast": round(contrast, 2),
"correlation": round(correlation, 3),
"energy": round(energy, 3),
"homogeneity": round(homogeneity, 3)
}
}
def _classify_texture(self, contrast, correlation, energy, homogeneity):
"""
Classify texture based on GLCM features
Returns (classification, description) tuple
"""
# Rule-based classification
if contrast < 100 and homogeneity > 0.8:
classification = "uniform and smooth"
description = "Model focused on a region with smooth, uniform texture showing consistent patterns with minimal variation"
elif correlation > 0.7 and energy > 0.3:
classification = "structured and regular"
description = "Model focused on a region with organized, structured patterns exhibiting regular, repeating elements"
elif contrast > 400 and correlation < 0.4:
classification = "irregular and chaotic"
description = "Model focused on a region with irregular, chaotic texture displaying highly variable patterns with no clear organization"
elif homogeneity < 0.5:
classification = "rough and coarse"
description = "Model focused on a region with rough, coarse texture showing sharp intensity changes and abrupt transitions"
elif energy < 0.2:
classification = "complex and varied"
description = "Model focused on a region with complex, varied texture containing multiple different patterns and high visual diversity"
else:
classification = "moderate texture"
description = "Model focused on a region with moderate texture complexity showing intermediate characteristics"
return classification, description
import warnings
warnings.filterwarnings("ignore")
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as transforms
from PIL import Image
import os
import numpy as np
import pandas as pd
from transformers import Dinov2Model
import openslide
from tqdm import tqdm
import logging
import matplotlib
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
from matplotlib.gridspec import GridSpec, GridSpecFromSubplotSpec
from openai import OpenAI
import time
logging.basicConfig(level=logging.INFO)
MODEL_PATH = "/kaggle/input/models/ulimaank/updated-diagnostic-model-jan-18/other/default/1/phase3_mil_best.pth"
OUTPUT_DIR = "/kaggle/working"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Device: {device}, GPUs: {torch.cuda.device_count()}")
DISEASE_NAMES = [
'Breast_cancer', 'annrbc-anemia_processed', 'colon_processed',
'leukemia_processed', 'lung_processed', 'oral-cancer_processed',
'ovarian-cancer_processed', 'sickle-cell-new_processed', 'thalassemia_processed'
]
STAGE_NAMES = {
0: 'Breast_cancer - ductal_carcinoma',
1: 'Breast_cancer - lobular_carcinoma',
2: 'Breast_cancer - mucinous_carcinoma',
3: 'Breast_cancer - papillary_carcinoma',
4: 'leukemia_processed - Early',
5: 'leukemia_processed - Pre',
6: 'leukemia_processed - Pro',
7: 'lung_processed - lung_aca',
8: 'lung_processed - lung_scc',
9: 'ovarian-cancer_processed - CC',
10: 'ovarian-cancer_processed - EC',
11: 'ovarian-cancer_processed - HGSC',
12: 'ovarian-cancer_processed - LGSC',
13: 'ovarian-cancer_processed - MC'
}
DISEASE_CLASS_MAPPING = {
0: "Breast_cancer",
1: "annrbc-anemia_processed",
2: "colon_processed",
3: "leukemia_processed",
4: "lung_processed",
5: "oral-cancer_processed",
6: "ovarian-cancer_processed",
7: "sickle-cell-new_processed",
8: "thalassemia_processed",
}
STAGE_CLASS_MAPPING = STAGE_NAMES
TARGET_SIZE = 256
standardize_transform = transforms.Resize((TARGET_SIZE, TARGET_SIZE))
# ================================================================
# DATA COLLECTION
# ================================================================
def collect_images_from_folder(folder_path):
images = []
valid_extensions = ('.svs', '.tif', '.ndpi', '.png', '.jpg', '.jpeg', '.tiff')
for root, dirs, files in os.walk(folder_path):
for f in files:
if f.lower().endswith(valid_extensions):
images.append(os.path.join(root, f))
return images
# ================================================================
# DATASET
# ================================================================
class SimpleSlideDataset(Dataset):
def __init__(self, image_paths, tile_size=224, max_tiles=1000):
self.image_paths = image_paths
self.tile_size = tile_size
self.max_tiles = max_tiles
def __len__(self):
return len(self.image_paths)
def __getitem__(self, idx):
slide_path = self.image_paths[idx]
tiles = []
try:
if slide_path.lower().endswith(('.svs', '.ndpi')):
slide = openslide.OpenSlide(slide_path)
width, height = slide.dimensions
for y in range(0, height, self.tile_size):
for x in range(0, width, self.tile_size):
if len(tiles) >= self.max_tiles:
break
tile = slide.read_region((x, y), 0, (self.tile_size, self.tile_size)).convert('RGB')
tiles.append(standardize_transform(tile))
if len(tiles) >= self.max_tiles:
break
slide.close()
elif slide_path.lower().endswith('.tif'):
try:
slide = openslide.OpenSlide(slide_path)
width, height = slide.dimensions
for y in range(0, height, self.tile_size):
for x in range(0, width, self.tile_size):
if len(tiles) >= self.max_tiles:
break
tile = slide.read_region((x, y), 0, (self.tile_size, self.tile_size)).convert('RGB')
tiles.append(standardize_transform(tile))
if len(tiles) >= self.max_tiles:
break
slide.close()
except openslide.OpenSlideError:
tiles = [standardize_transform(Image.open(slide_path).convert('RGB'))]
except Exception:
tiles = [standardize_transform(Image.open(slide_path).convert('RGB'))]
else:
tiles = [standardize_transform(Image.open(slide_path).convert('RGB'))]
if not tiles:
raise ValueError("No tiles extracted")
return tiles, slide_path
except Exception as e:
logging.error(f"Error processing slide {slide_path}: {e}")
return [], slide_path
test_transform = transforms.Compose([
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
def simple_collate(batch):
valid_batch = [item for item in batch if item[0]]
if not valid_batch:
return [], []
tiles_list, paths = zip(*valid_batch)
processed_tiles = [torch.stack([test_transform(tile) for tile in tiles]) for tiles in tiles_list]
return processed_tiles, list(paths)
# ================================================================
# MODEL ARCHITECTURE
# ================================================================
class ViTBackbone(nn.Module):
def __init__(self):
super().__init__()
self.vit = Dinov2Model.from_pretrained("owkin/phikon-v2")
def forward(self, x):
return self.vit(pixel_values=x).last_hidden_state[:, 0]
class ClassificationHead(nn.Module):
def __init__(self, in_dim=1024, num_classes=2, hidden_dim=512):
super().__init__()
self.classifier = nn.Sequential(
nn.Linear(in_dim, hidden_dim),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(hidden_dim, num_classes)
)
def forward(self, x):
return self.classifier(x)
class HierarchicalMILAggregator(nn.Module):
def __init__(self, embed_dim=1024, num_heads=8, num_layers=2,
num_diseases=6, num_stage_classes=0, disease_names=None):
super().__init__()
self.pre_norm = nn.LayerNorm(embed_dim)
encoder_layer = nn.TransformerEncoderLayer(
d_model=embed_dim, nhead=num_heads, batch_first=True, dropout=0.1)
self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
self.attention = nn.Sequential(nn.Linear(embed_dim, 256), nn.Tanh(), nn.Linear(256, 1))
self.disease_head = ClassificationHead(embed_dim, num_diseases)
self.severity_heads = nn.ModuleDict()
for name in disease_names:
self.severity_heads[name] = ClassificationHead(embed_dim, 2)
self.stage_head = ClassificationHead(embed_dim, num_stage_classes) if num_stage_classes > 0 else None
self.disease_name_to_idx = {n: i for i, n in enumerate(disease_names)}
self.idx_to_disease_name = {i: n for n, i in self.disease_name_to_idx.items()}
self.disease_names = disease_names
def forward(self, tile_features):
normalized = self.pre_norm(tile_features)
aggregated = self.transformer(normalized)
attn_scores = self.attention(aggregated)
attn_weights = torch.softmax(attn_scores.squeeze(-1), dim=1)
weighted = torch.sum(aggregated * attn_weights.unsqueeze(-1), dim=1)
disease_logits = self.disease_head(weighted)
severity_logits = {n: self.severity_heads[n](weighted) for n in self.disease_names}
stage_logits = self.stage_head(weighted) if self.stage_head is not None else None
return disease_logits, severity_logits, stage_logits, attn_weights
class Phase3Model(nn.Module):
def __init__(self, backbone, num_diseases=6, num_stage_classes=0, disease_names=None):
super().__init__()
self.backbone = backbone
for param in self.backbone.parameters():
param.requires_grad = False
self.aggregator = HierarchicalMILAggregator(
num_diseases=num_diseases,
num_stage_classes=num_stage_classes,
disease_names=disease_names
)
def forward(self, tiles, enable_gradients=False):
all_features = []
for batch_tiles in tiles:
if batch_tiles.numel() == 0:
continue
batch_tiles = batch_tiles.to(next(self.backbone.parameters()).device)
if enable_gradients:
batch_features = self.backbone(batch_tiles)
else:
with torch.no_grad():
batch_features = self.backbone(batch_tiles)
all_features.append(batch_features)
if not all_features:
raise ValueError("No valid tile features could be extracted.")
all_features = torch.stack(all_features)
return self.aggregator(all_features)
# ================================================================
# PREDICTION
# ================================================================
def predict_image(model, tiles, disease_names, stage_names):
model.eval()
try:
with torch.no_grad():
disease_logits, severity_logits, stage_logits, _ = model(tiles)
disease_probs = F.softmax(disease_logits, dim=1)
disease_pred_idx = torch.argmax(disease_probs, dim=1).item()
disease_confidence = disease_probs[0, disease_pred_idx].item()
predicted_disease = disease_names[disease_pred_idx]
severity_probs = F.softmax(severity_logits[predicted_disease], dim=1)
severity_pred = torch.argmax(severity_probs, dim=1).item()
severity_confidence = severity_probs[0, severity_pred].item()
severity_label = "Normal" if severity_pred == 0 else "Abnormal"
stage_label = "N/A"
stage_confidence = 0.0
if severity_pred == 1 and stage_logits is not None:
stage_probs = F.softmax(stage_logits, dim=1)
stage_pred_idx = torch.argmax(stage_probs, dim=1).item()
stage_confidence = stage_probs[0, stage_pred_idx].item()
stage_label = stage_names.get(stage_pred_idx, f"Stage_{stage_pred_idx}")
return {
'disease': predicted_disease,
'disease_confidence': disease_confidence,
'severity': severity_label,
'severity_confidence': severity_confidence,
'stage': stage_label,
'stage_confidence': stage_confidence
}
except Exception as e:
logging.error(f"Error during prediction: {e}")
return None
# ================================================================
# GPT EXPLANATION GENERATOR
# ================================================================
def generate_comprehensive_explanation(comprehensive_data):
"""
Calls GPT-4o-mini to convert technical XAI metrics into a human-friendly
explanation. Falls back to a template string if the API call fails.
Only references Attention and GradCAM heatmaps.
"""
try:
client = OpenAI(api_key=os.getenv('OPENAI_API_KEY'))
prompt = f"""You are an AI explainability assistant helping users understand how a hierarchical medical image classification model made its decision. Convert the following technical analysis into a clear, accessible explanation.
HIERARCHICAL MODEL PREDICTION:
- Region: {comprehensive_data['predicted_disease']} ({comprehensive_data['gradcam_disease_conf']:.1%} confidence)
- Status Level: {comprehensive_data['gradcam_severity']} ({comprehensive_data['gradcam_severity_conf']:.1%} confidence)
- Stage Level: {comprehensive_data['predicted_stage']} ({comprehensive_data['stage_confidence']:.1%} confidence)
GRADCAM ANALYSIS (Gradient-weighted Class Activation Mapping):
- Note: Bright/warm regions in GradCAM indicate areas that most strongly influenced the model's prediction
SPATIAL ATTENTION PATTERN AND VISUAL CHARACTERISTICS (from Attention Heatmap):
- Primary Focus: {comprehensive_data['primary_position']} (intensity: {comprehensive_data['primary_intensity']:.2f})
- Attention Hotspots: {comprehensive_data['hotspot_count']}
- Spatial Distribution: Center {comprehensive_data['center_attention']:.1f}%, Mid-region {comprehensive_data['mid_attention']:.1f}%, Periphery {comprehensive_data['periphery_attention']:.1f}%
- Clustering: {comprehensive_data['scatter_level']} scatter level with {comprehensive_data['num_clusters']} clusters
- Dominant Color: {comprehensive_data['dominant_color']} ({comprehensive_data['color_confidence']:.1f}% confidence)
- Texture Pattern: {comprehensive_data['texture_classification']}
- Texture Scores: Uniformity {comprehensive_data['uniformity']}/100, Organization {comprehensive_data['organization']}/100, Complexity {comprehensive_data['complexity']}/100, Smoothness {comprehensive_data['smoothness']}/100
CRITICAL INSTRUCTIONS:
1. Write in clear, accessible language for someone without medical or technical expertise
2. Ground ALL statements in the provided data - do NOT add medical interpretations or diagnoses
3. Explain how the two explainability methods (Attention Heatmap and GradCAM) show WHERE the model focused
4. Describe WHAT visual patterns were detected, not WHY medically
5. Keep it concise but informative (under 100 words)
6. Structure with clear sections
7. Make it conversational but professional
8. Visual Characteristics and Spatial Attention Pattern were taken from Attention Heatmap
Generate a comprehensive explanation covering: what the model decided, where it looked, what the Attention and GradCAM methods revealed, what visual characteristics were important, and how confident we can be in the decision.
Format as natural paragraphs, not bullet points."""
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": "You are an expert at explaining complex AI model decisions in simple, clear language. You help users understand model behavior without making medical claims."},
{"role": "user", "content": prompt}
],
temperature=0.7,
max_tokens=700
)
return response.choices[0].message.content.strip()
except Exception as e:
logging.warning(f"OpenAI API call failed: {e}. Using fallback template.")
return (
f"MODEL DECISION SUMMARY\n\n"
f"The model classified this as '{comprehensive_data['predicted_disease']}' "
f"with severity '{comprehensive_data['gradcam_severity']}' "
f"and stage '{comprehensive_data['predicted_stage']}' "
f"({comprehensive_data['stage_confidence']:.1%} confidence).\n\n"
f"ATTENTION ANALYSIS\n\n"
f"Primary focus: {comprehensive_data['primary_position']} region. "
f"Attention shows {comprehensive_data['scatter_level']} scatter across "
f"{comprehensive_data['num_clusters']} clusters. "
f"Distribution - Center: {comprehensive_data['center_attention']:.1f}%, "
f"Mid: {comprehensive_data['mid_attention']:.1f}%, "
f"Periphery: {comprehensive_data['periphery_attention']:.1f}%.\n\n"
f"GRADCAM ANALYSIS\n\n"
f"GradCAM confidence: {comprehensive_data['gradcam_disease_conf']:.1%} (disease), "
f"{comprehensive_data['gradcam_severity_conf']:.1%} (severity).\n\n"
f"VISUAL PATTERNS\n\n"
f"Dominant color: {comprehensive_data['dominant_color']}. "
f"Texture: {comprehensive_data['texture_classification']} "
f"(uniformity {comprehensive_data['uniformity']}/100, "
f"smoothness {comprehensive_data['smoothness']}/100)."
)
# ================================================================
# RENDER EXPLANATION TEXT -> RGB NUMPY ARRAY FOR ROW 3
# ================================================================
def _render_explanation_to_image(explanation_text, figsize=(16, 4)):
"""
Renders a plain-text explanation string into an (H, W, 3) uint8 numpy
array that fills the entire Row 3 panel in display_prediction().
"""
fig, ax = plt.subplots(figsize=figsize, facecolor='#0F0F2A')
fig.subplots_adjust(left=0, right=1, top=1, bottom=0)
ax.set_facecolor('#0F0F2A')
ax.set_xlim(0, 1)
ax.set_ylim(0, 1)
ax.axis('off')
ax.add_patch(mpatches.FancyBboxPatch(
(0.0, 0.0), 1.0, 1.0,
boxstyle="round,pad=0.01",
linewidth=2,
edgecolor='#3498DB',
facecolor='#16213E',
transform=ax.transAxes,
clip_on=False
))
ax.text(
0.5, 0.93,
'Textual Explanation',
ha='center', va='top',
fontsize=11,
fontweight='bold',
color='#3498DB',
transform=ax.transAxes
)
ax.add_line(plt.Line2D(
[0.02, 0.98], [0.855, 0.855],
transform=ax.transAxes,
color='#3498DB',
linewidth=1.0
))
ax.text(
0.02, 0.83,
explanation_text,
va='top', ha='left',
fontsize=10,
color='#E0E0F0',
family='monospace',
wrap=True,
transform=ax.transAxes
)
fig.canvas.draw()
buf = fig.canvas.buffer_rgba()
img_array = np.frombuffer(buf, dtype=np.uint8).reshape(
fig.canvas.get_width_height()[::-1] + (4,)
)
plt.close(fig)
return img_array[:, :, :3]
# ================================================================
# ATTENTION OVERLAY HELPERS
# ================================================================
def _build_attention_overlay(img_array, heatmap_raw):
"""
Takes img_array (H, W, 3) uint8 and heatmap_raw (H, W) float,
returns an RGB overlay as np.ndarray in [0, 1].
"""
import cv2
img_norm = img_array.astype(np.float32) / 255.0
hm = heatmap_raw.astype(np.float32)
h, w = img_norm.shape[:2]
if hm.shape != (h, w):
hm = cv2.resize(hm, (w, h), interpolation=cv2.INTER_CUBIC)
hm_min, hm_max = hm.min(), hm.max()
if hm_max > hm_min:
hm = (hm - hm_min) / (hm_max - hm_min)
cmap = matplotlib.colormaps.get_cmap('jet')
hm_colored = cmap(hm)[:, :, :3]
overlay = img_norm * 0.5 + hm_colored * 0.5
return np.clip(overlay, 0, 1)
def _preprocess_images_for_attention(image_paths):
processed = []
for p in image_paths:
try:
img_pil = Image.open(p).convert('RGB').resize((224, 224), Image.BILINEAR)
tensor = test_transform(img_pil)
processed.append(tensor)
except Exception as e:
logging.warning(f"Could not preprocess {p} for attention: {e}")
processed.append(torch.zeros(3, 224, 224))
return processed
# ================================================================
# DISPLAY FUNCTION - 3-ROW LAYOUT
# Row 1 : Original Image | Diagnostic Report
# Row 2 : Attention Heatmap | GradCAM Heatmap
# Row 3 : GPT-Generated Human-Friendly Text Explanation
# ================================================================
def display_prediction(image_path, prediction,
heatmap_images=None,
heatmap_titles=None,
explanation_image=None):
severity = prediction['severity']
accent = '#E74C3C' if severity == 'Abnormal' else '#2ECC71'
bg_color = '#1A1A2E'
panel_color = '#16213E'
border_dim = '#2A2A4A'
# Exactly 2 heatmaps: Attention + GradCAM
if heatmap_titles is None:
heatmap_titles = ['Attention Heatmap', 'GradCAM Heatmap']
fig = plt.figure(figsize=(16, 14), facecolor=bg_color)
outer_gs = GridSpec(3, 1, figure=fig,
height_ratios=[5, 4, 3],
hspace=0.08)
# ===== ROW 1: Original Image | Diagnostic Report =====
row1_gs = GridSpecFromSubplotSpec(1, 2,
subplot_spec=outer_gs[0],
width_ratios=[1, 1.2],
wspace=0.05)
ax_img = fig.add_subplot(row1_gs[0])
ax_img.set_facecolor(bg_color)
try:
img = Image.open(image_path).convert('RGB')
ax_img.imshow(img)
except Exception:
ax_img.text(0.5, 0.5, 'WSI / Slide\n(preview unavailable)',
ha='center', va='center', color='white', fontsize=13,
transform=ax_img.transAxes)
for spine in ax_img.spines.values():
spine.set_edgecolor(accent)
spine.set_linewidth(3)
ax_img.set_xticks([])
ax_img.set_yticks([])
ax_img.set_title(os.path.basename(image_path), color='white',
fontsize=11, pad=8, fontweight='bold')
ax_info = fig.add_subplot(row1_gs[1])
ax_info.set_facecolor(bg_color)
ax_info.set_xlim(0, 1)
ax_info.set_ylim(0, 1)
ax_info.axis('off')
ax_info.text(0.5, 0.96, 'Diagnostic Report',
ha='center', va='top', fontsize=15, fontweight='bold',
color='white', transform=ax_info.transAxes)
divider = plt.Line2D([0.05, 0.95], [0.89, 0.89],
transform=ax_info.transAxes,
color=accent, linewidth=1.5)
ax_info.add_line(divider)
def draw_card(ax, y, label, value, confidence, color):
ax.add_patch(mpatches.FancyBboxPatch(
(0.04, y - 0.11), 0.92, 0.14,
boxstyle="round,pad=0.01",
linewidth=1.5, edgecolor=color,
facecolor=panel_color, transform=ax.transAxes, clip_on=False
))
ax.text(0.10, y - 0.01, label.upper(),
ha='left', va='center', fontsize=8, color='#A0A0C0',
fontweight='bold', transform=ax.transAxes)
ax.text(0.10, y - 0.05, value,
ha='left', va='center', fontsize=13, color='white',
fontweight='bold', transform=ax.transAxes)
if confidence > 0:
bar_y = y - 0.09
ax.add_patch(mpatches.FancyBboxPatch(
(0.08, bar_y), 0.60, 0.015,
boxstyle="round,pad=0.001", linewidth=0,
facecolor='#0F3460', transform=ax.transAxes, clip_on=False
))
ax.add_patch(mpatches.FancyBboxPatch(
(0.08, bar_y), 0.60 * confidence, 0.015,
boxstyle="round,pad=0.001", linewidth=0,
facecolor=color, transform=ax.transAxes, clip_on=False
))
ax.text(0.72, bar_y + 0.007, f'{confidence:.1%}',
ha='left', va='center', fontsize=9, color=color,
fontweight='bold', transform=ax.transAxes)
draw_card(ax_info, 0.78,
'Region',
prediction['disease'].replace('_processed', '').replace('_', ' ').title(),
prediction['disease_confidence'], '#3498DB')
draw_card(ax_info, 0.57, 'Status', severity,
prediction['severity_confidence'], accent)
stage_val = prediction['stage']
stage_conf = prediction['stage_confidence']
if stage_val == 'N/A':
stage_display = 'N/A (Normal)'
stage_conf = 0
else:
stage_display = (stage_val.split(' - ')[-1].replace('_', ' ').title()
if ' - ' in stage_val else stage_val)
draw_card(ax_info, 0.36, 'Stage / Subtype', stage_display, stage_conf, '#F39C12')
# ===== ROW 2: Attention Heatmap | GradCAM Heatmap (2 columns) =====
row2_gs = GridSpecFromSubplotSpec(1, 2,
subplot_spec=outer_gs[1],
wspace=0.06)
for col_idx in range(2):
ax_hm = fig.add_subplot(row2_gs[col_idx])
ax_hm.set_facecolor(panel_color)
if heatmap_images and col_idx < len(heatmap_images) and heatmap_images[col_idx] is not None:
hm = heatmap_images[col_idx]
ax_hm.imshow(hm if isinstance(hm, np.ndarray) else np.array(hm))
else:
ax_hm.set_xlim(0, 1)
ax_hm.set_ylim(0, 1)
ax_hm.add_patch(mpatches.FancyBboxPatch(
(0.05, 0.05), 0.90, 0.90,
boxstyle="round,pad=0.02",
linewidth=1.5, linestyle='--',
edgecolor='#4A4A6A', facecolor='#0F0F2A',
transform=ax_hm.transAxes, clip_on=False
))
ax_hm.text(0.5, 0.5, '[ Heatmap\nPlaceholder ]',
ha='center', va='center',
color='#4A4A6A', fontsize=9, fontstyle='italic',
transform=ax_hm.transAxes)
title = heatmap_titles[col_idx] if col_idx < len(heatmap_titles) else f'Heatmap {col_idx+1}'
ax_hm.set_title(title, color='#A0A0C0', fontsize=9, fontweight='bold', pad=5)
for spine in ax_hm.spines.values():
spine.set_edgecolor(border_dim)
spine.set_linewidth(1.2)
ax_hm.set_xticks([])
ax_hm.set_yticks([])
# ===== ROW 3: GPT Explanation =====
ax_text = fig.add_subplot(outer_gs[2])
ax_text.set_facecolor(panel_color)
if explanation_image is not None:
exp_img = explanation_image if isinstance(explanation_image, np.ndarray) \
else np.array(explanation_image)
ax_text.imshow(exp_img, aspect='auto')
ax_text.set_xticks([])
ax_text.set_yticks([])
else:
ax_text.set_xlim(0, 1)
ax_text.set_ylim(0, 1)
ax_text.axis('off')
ax_text.add_patch(mpatches.FancyBboxPatch(
(0.01, 0.05), 0.98, 0.90,
boxstyle="round,pad=0.02",
linewidth=1.5, linestyle='--',
edgecolor='#4A4A6A', facecolor='#0F0F2A',
transform=ax_text.transAxes, clip_on=False
))
ax_text.text(0.5, 0.80,
'Human-Friendly Text Explanation',
ha='center', va='center',
color='#A0A0C0', fontsize=11, fontweight='bold',
transform=ax_text.transAxes)
ax_text.text(0.5, 0.42,
'[ Textual Explanation Placeholder ]\n\n'
'The model focused on [extracted features]\n'
'because [rule-based reasoning] ...',
ha='center', va='center',
color='#3A3A5A', fontsize=10, fontstyle='italic',
transform=ax_text.transAxes)
for spine in ax_text.spines.values():
spine.set_edgecolor(border_dim)
spine.set_linewidth(1.2)
plt.tight_layout(pad=1.2)
plt.show()
print()
# ================================================================
# MODEL LOADER
# ================================================================
def load_model():
if not os.path.exists(MODEL_PATH):
print(f"ERROR: Model not found at {MODEL_PATH}")
return None, None
checkpoint = torch.load(MODEL_PATH, map_location=device)
state_dict = checkpoint['model_state_dict']
if list(state_dict.keys())[0].startswith('module.'):
state_dict = {k.replace('module.', ''): v for k, v in state_dict.items()}
severity_head_names = set()
for k in state_dict.keys():
if k.startswith('aggregator.severity_heads.'):
parts = k.split('.')
if len(parts) > 2:
severity_head_names.add(parts[2])
disease_names = sorted(list(severity_head_names))
num_diseases = state_dict['aggregator.disease_head.classifier.3.weight'].shape[0]
num_stage_classes = (
state_dict['aggregator.stage_head.classifier.3.weight'].shape[0]
if 'aggregator.stage_head.classifier.3.weight' in state_dict else 0
)
print(f"\nModel config -> diseases: {num_diseases} | stages: {num_stage_classes}")
print(f"Classes: {', '.join(disease_names)}\n")
backbone = ViTBackbone()
model = Phase3Model(backbone, num_diseases=num_diseases,
num_stage_classes=num_stage_classes,
disease_names=disease_names).to(device)
model.load_state_dict(state_dict, strict=True)
model.eval()
print("Model loaded successfully!\n")
return model, disease_names
# ================================================================
# MAIN INFERENCE PIPELINE
# ================================================================
def run_inference():
print("\n" + "=" * 70)
print(" HIERARCHICAL MIL MODEL - PATHOLOGY INFERENCE")
print("=" * 70)
print("\nOptions:")
print(" 1. Single image (provide full path to one image file)")
print(" 2. Folder (provide path to a folder; all images processed)")
choice = input("\nSelect option (1 / 2): ").strip()
if choice == '1':
image_path = input("Enter image path: ").strip()
if not os.path.isfile(image_path):
print(f"ERROR: File not found -> {image_path}")
return
all_images = [image_path]
elif choice == '2':
folder_path = input("Enter folder path: ").strip()
if not os.path.isdir(folder_path):
print(f"ERROR: Folder not found -> {folder_path}")
return
all_images = collect_images_from_folder(folder_path)
if not all_images:
print("No valid images found in the folder.")
return
print(f"Found {len(all_images)} image(s).")
else:
print("Invalid option.")
return
# ----------------------------------------------------------
# Load model
# ----------------------------------------------------------
print("\nLoading model ...")
model, disease_names = load_model()
if model is None:
return
# ----------------------------------------------------------
# STEP 1: Run Attention and GradCAM analyses only
# ----------------------------------------------------------
print("\n" + "=" * 70)
print("Running Attention Analysis ...")
print("=" * 70)
processed_images = _preprocess_images_for_attention(all_images)
attn_results = run_attention_analysis(
attention_extractor,
device,
all_images,
processed_images,
DISEASE_CLASS_MAPPING,
STAGE_CLASS_MAPPING
)
print("\n" + "=" * 70)
print("Running GradCAM Analysis ...")
print("=" * 70)
gradcam_results = run_tri_head_gradcam_plus_plus_analysis(model, device, all_images)
# ----------------------------------------------------------
# STEP 2: Feature extraction from attention heatmaps
# ----------------------------------------------------------
print("\n" + "=" * 70)
print("Running Feature Extraction ...")
print("=" * 70)
explanations_list = []
for i, attention_result in enumerate(attn_results):
print(f" [{i+1}/{len(attn_results)}] Extracting features: {attention_result['filename']}")
extractor = HeatmapFeatureExtractor(attention_result)
bright = extractor.get_brightest_region()
scatter = extractor.get_activation_scatter()
dom_color = extractor.get_dominant_focus_color()
texture = extractor.get_texture_analysis()
explanations_list.append({
"brightest": bright,
"scatter": scatter,
"dominant_color": dom_color,
"texture": texture,
})
print(f" ✅ Position={bright['primary_hotspot']['position']}, "
f"Scatter={scatter['scatter_level']}, "
f"Color={dom_color['dominant_color_name']}, "
f"Texture={texture['texture_classification']}")
# ----------------------------------------------------------
# STEP 3: Standard inference loop + GPT explanation + display
# ----------------------------------------------------------
dataset = SimpleSlideDataset(all_images)
dataloader = DataLoader(dataset, batch_size=1, shuffle=False,
collate_fn=simple_collate, num_workers=2, pin_memory=True)
results = []
for batch_idx, batch in enumerate(tqdm(dataloader, desc="Running inference")):
tiles, paths = batch
if not tiles or not paths:
continue
slide_path = paths[0]
try:
img_idx = all_images.index(slide_path)
except ValueError:
img_idx = batch_idx
prediction = predict_image(model, tiles, disease_names, STAGE_NAMES)
if prediction is None:
print(f"Failed to process: {slide_path}")
continue
# ---- Build the 2 overlay images: Attention + GradCAM ----
attn_overlay = None
if attn_results and img_idx < len(attn_results):
ar = attn_results[img_idx]
attn_overlay = _build_attention_overlay(ar['image'], ar['attention_heatmap'])
gradcam_overlay = None
if gradcam_results and img_idx < len(gradcam_results):
gradcam_overlay = gradcam_results[img_idx]['union']['heatmap_overlay']
# ---- Build comprehensive_data dict for GPT ----
exp = explanations_list[img_idx]
bright = exp['brightest']
scatter_res = exp['scatter']
dom = exp['dominant_color']
texture = exp['texture']
# GradCAM-sourced confidence and severity values
gradcam_disease_conf = gradcam_results[img_idx]['level1_disease']['confidence'] \
if gradcam_results else 0.0
gradcam_severity_conf = gradcam_results[img_idx]['level2_severity']['confidence'] \
if gradcam_results else 0.0
gradcam_severity = gradcam_results[img_idx]['level2_severity']['predicted_class'] \
if gradcam_results else 'N/A'
comprehensive_data = {
'predicted_disease': prediction['disease'],
'gradcam_severity': gradcam_severity,
'predicted_stage': prediction['stage'],
'stage_confidence': prediction['stage_confidence'],
'gradcam_disease_conf': gradcam_disease_conf,
'gradcam_severity_conf': gradcam_severity_conf,
'primary_position': bright['primary_hotspot']['position'],
'primary_intensity': bright['primary_hotspot']['intensity'],
'hotspot_count': bright['hotspot_count'],
'center_attention': bright['spatial_coverage']['center_attention'],
'mid_attention': bright['spatial_coverage']['mid_region_attention'],
'periphery_attention': bright['spatial_coverage']['periphery_attention'],
'scatter_level': scatter_res['scatter_level'],
'num_clusters': scatter_res['num_clusters'],
'dominant_color': dom['dominant_color_name'],
'color_confidence': dom['color_confidence'],
'texture_classification': texture['texture_classification'],
'uniformity': texture['texture_scores']['uniformity'],
'organization': texture['texture_scores']['organization'],
'complexity': texture['texture_scores']['complexity'],
'smoothness': texture['texture_scores']['smoothness'],
}
# ---- Generate GPT explanation -> render to image for Row 3 ----
print(f"\n Generating GPT explanation for image {img_idx + 1} ...")
explanation_text = generate_comprehensive_explanation(comprehensive_data)
explanation_image = _render_explanation_to_image(explanation_text)
time.sleep(0.5)
# ---- Display: Original | Report | Attention | GradCAM | Explanation ----
display_prediction(
image_path = slide_path,
prediction = prediction,
heatmap_images = [attn_overlay, gradcam_overlay],
heatmap_titles = ['Attention Heatmap', 'GradCAM Heatmap'],
explanation_image = explanation_image
)
results.append({
'image_path': slide_path,
'image_name': os.path.basename(slide_path),
'predicted_disease': prediction['disease'],
'disease_confidence': prediction['disease_confidence'],
'predicted_severity': prediction['severity'],
'severity_confidence': prediction['severity_confidence'],
'predicted_stage': prediction['stage'],
'stage_confidence': prediction['stage_confidence']
})
if results:
df = pd.DataFrame(results)
out_path = os.path.join(OUTPUT_DIR, "inference_results.csv")
df.to_csv(out_path, index=False)
print(f"\nResults saved -> {out_path}")
print(f"Total processed: {len(results)}")
print("\n" + "=" * 70)
print("DONE")
print("=" * 70)
if __name__ == "__main__":
run_inference() |