File size: 33,859 Bytes
86aa283
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
RadioFlow: AI-Powered Radiology Workflow Agent
Kaggle Notebook with REAL MedGemma Model
MedGemma Impact Challenge Submission
"""

# %% [markdown]
# # 🩻 RadioFlow: AI-Powered Radiology Workflow Agent
# ## MedGemma Impact Challenge Submission
#
# **Author:** Samarpeet Garad
# **Date:** February 2026
#
# ---
#
# ## Executive Summary
#
# RadioFlow is a **real AI-powered** multi-agent system that analyzes chest X-rays using
# Google's **MedGemma** model. This notebook runs with actual model inference on Kaggle's
# free GPU, demonstrating production-ready medical AI.
#
# **Key Features:**
# - πŸ€– Real MedGemma-4B model inference (not simulated!)
# - πŸ”¬ 4-agent orchestrated pipeline
# - πŸ“‹ Generates structured radiology reports
# - 🚦 Automatic priority assessment and routing

# %% [markdown]
# ## 1. Setup and GPU Check

# %%
import os
import sys
import time
import json
import warnings

warnings.filterwarnings("ignore")

# Check GPU availability
import torch

print(f"PyTorch version: {torch.__version__}")
print(f"CUDA available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    print(
        f"GPU Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB"
    )
else:
    print("⚠️ No GPU detected - model will run slower on CPU")

# %%
# Install required packages
print("πŸ“¦ Installing required packages...")
import subprocess

subprocess.run(
    ["pip", "install", "-q", "bitsandbytes", "accelerate", "sentencepiece"], check=True
)
print("βœ… Packages installed!")

# %%
import numpy as np
import pandas as pd
from datetime import datetime
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Any, Tuple
from PIL import Image, ImageDraw
import matplotlib.pyplot as plt

# Hugging Face
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from huggingface_hub import login

# Display
from IPython.display import HTML, display, Markdown, clear_output
import plotly.graph_objects as go

print("βœ… Dependencies loaded successfully")

# %% [markdown]
# ## 2. Authenticate with Hugging Face
#
# To use MedGemma, you need to:
# 1. Accept the license at https://huggingface.co/google/medgemma-4b-it
# 2. Add your HF token as a Kaggle secret named "HF_TOKEN"

# %%
# Get HuggingFace token from Kaggle secrets
try:
    from kaggle_secrets import UserSecretsClient

    secrets = UserSecretsClient()
    HF_TOKEN = secrets.get_secret("HF_TOKEN")
    login(token=HF_TOKEN)
    print("βœ… Authenticated with Hugging Face")
except Exception as e:
    print(f"⚠️ Could not get HF token from Kaggle secrets: {e}")
    print("Please add your HF_TOKEN as a Kaggle secret")
    HF_TOKEN = None

# %% [markdown]
# ## 3. Load Real MedGemma Model

# %%
MODEL_NAME = "google/medgemma-4b-it"

# Configure 4-bit quantization for efficient memory usage
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)

print(f"πŸ”„ Loading {MODEL_NAME}...")
print("   This may take 1-2 minutes on first run...")

try:
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        quantization_config=quantization_config,
        device_map="auto",
        trust_remote_code=True,
        torch_dtype=torch.float16,
    )
    model.eval()

    MODEL_LOADED = True
    print(f"βœ… MedGemma loaded successfully!")
    print(f"   Memory used: {torch.cuda.memory_allocated() / 1e9:.2f} GB")

except Exception as e:
    print(f"❌ Failed to load model: {e}")
    print("   Falling back to demo mode...")
    MODEL_LOADED = False
    model = None
    tokenizer = None


# %%
def generate_medgemma_response(prompt: str, max_tokens: int = 512) -> str:
    """Generate response using real MedGemma model."""
    if not MODEL_LOADED:
        return "[Demo mode - model not loaded]"

    messages = [{"role": "user", "content": prompt}]

    # Tokenize with proper attention mask
    inputs = tokenizer.apply_chat_template(
        messages, return_tensors="pt", add_generation_prompt=True
    )

    # Create attention mask (1 for all tokens since no padding)
    attention_mask = torch.ones_like(inputs)

    # Move to device
    inputs = inputs.to(model.device)
    attention_mask = attention_mask.to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            inputs,
            attention_mask=attention_mask,
            max_new_tokens=max_tokens,
            do_sample=False,  # Use greedy decoding to avoid numerical issues
            pad_token_id=tokenizer.eos_token_id,
        )

    response = tokenizer.decode(outputs[0][inputs.shape[1] :], skip_special_tokens=True)
    return response.strip()


# Test the model
if MODEL_LOADED:
    print("\nπŸ§ͺ Testing MedGemma...")
    test_response = generate_medgemma_response(
        "What are the key findings to look for in a chest X-ray? List 3 briefly.",
        max_tokens=100,
    )
    print(f"Response: {test_response[:200]}...")

# %% [markdown]
# ## 4. Agent Architecture
#
# RadioFlow uses a 4-agent pipeline, each powered by MedGemma:
#
# ```
# β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”    β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”    β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”    β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
# β”‚  CXR Analyzer  │───▢│    Finding     │───▢│    Report      │───▢│   Priority     β”‚
# β”‚ (Image Analysisβ”‚    β”‚  Interpreter   β”‚    β”‚   Generator    β”‚    β”‚    Router      β”‚
# β”‚   + MedGemma)  β”‚    β”‚  (MedGemma)    β”‚    β”‚  (MedGemma)    β”‚    β”‚  (MedGemma)    β”‚
# β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜    β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜    β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜    β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
# ```


# %%
@dataclass
class AgentResult:
    """Standardized result from any agent"""

    agent_name: str
    status: str
    data: Dict[str, Any]
    processing_time_ms: float
    timestamp: str = field(default_factory=lambda: datetime.now().isoformat())


class BaseAgent:
    """Base class for all RadioFlow agents"""

    def __init__(self, name: str, model_name: str):
        self.name = name
        self.model_name = model_name

    def __call__(self, input_data: Any, context: Optional[Dict] = None) -> AgentResult:
        start = time.time()
        result = self.process(input_data, context)
        result.processing_time_ms = (time.time() - start) * 1000
        return result

    def process(self, input_data: Any, context: Optional[Dict] = None) -> AgentResult:
        raise NotImplementedError


print("βœ… Base agent class defined")

# %% [markdown]
# ## 5. Agent Implementations with Real MedGemma


# %%
class CXRAnalyzerAgent(BaseAgent):
    """
    Agent 1: Image Analyzer
    Analyzes chest X-ray images using computer vision + MedGemma.
    """

    def __init__(self):
        super().__init__("CXR Analyzer", "MedGemma + Image Analysis")
        self.regions = [
            "right_upper_lung",
            "right_middle_lung",
            "right_lower_lung",
            "left_upper_lung",
            "left_lower_lung",
            "cardiac_silhouette",
            "mediastinum",
            "costophrenic_angles",
        ]

    def process(
        self, image: Image.Image, context: Optional[Dict] = None
    ) -> AgentResult:
        # Analyze image characteristics
        img_array = np.array(image.convert("L"))  # Grayscale

        # Calculate regional statistics
        h, w = img_array.shape
        regions_stats = {
            "right_lung": img_array[:, w // 2 :].mean(),
            "left_lung": img_array[:, : w // 2].mean(),
            "upper": img_array[: h // 2, :].mean(),
            "lower": img_array[h // 2 :, :].mean(),
            "cardiac": img_array[h // 3 : 2 * h // 3, w // 3 : 2 * w // 3].mean(),
        }

        overall_brightness = img_array.mean()
        contrast = img_array.std()
        asymmetry = abs(regions_stats["right_lung"] - regions_stats["left_lung"])

        # Generate findings based on image analysis
        findings = []

        # Check for opacities (darker regions than expected)
        if regions_stats["lower"] > overall_brightness + 10:
            findings.append(
                {
                    "type": "opacity",
                    "region": "lower_lung_zones",
                    "confidence": min(0.95, 0.7 + asymmetry / 50),
                    "severity": "moderate"
                    if regions_stats["lower"] > overall_brightness + 20
                    else "mild",
                    "description": f"Increased density in lower lung zones (mean: {regions_stats['lower']:.0f})",
                }
            )

        # Check for asymmetry
        if asymmetry > 15:
            side = (
                "right"
                if regions_stats["right_lung"] > regions_stats["left_lung"]
                else "left"
            )
            findings.append(
                {
                    "type": "asymmetry",
                    "region": f"{side}_hemithorax",
                    "confidence": min(0.9, 0.6 + asymmetry / 30),
                    "severity": "mild",
                    "description": f"Asymmetric density noted, {side} side appears denser",
                }
            )

        # Check cardiac region
        if regions_stats["cardiac"] > overall_brightness + 25:
            findings.append(
                {
                    "type": "cardiomegaly",
                    "region": "cardiac_silhouette",
                    "confidence": 0.75,
                    "severity": "mild",
                    "description": "Enlarged cardiac silhouette suggested",
                }
            )

        # If no abnormalities, report normal
        if not findings:
            findings.append(
                {
                    "type": "normal",
                    "region": "bilateral_lungs",
                    "confidence": 0.85,
                    "severity": "none",
                    "description": "No significant abnormalities detected on initial analysis",
                }
            )

        # Use MedGemma to enhance the analysis
        if MODEL_LOADED and findings:
            finding_desc = "; ".join([f["description"] for f in findings])
            enhancement_prompt = f"""As a radiologist, given these image analysis findings:
{finding_desc}

Provide a brief (2-3 sentence) clinical interpretation of what these findings might indicate.
Focus on clinical relevance."""

            enhanced = generate_medgemma_response(enhancement_prompt, max_tokens=100)
            clinical_note = enhanced
        else:
            clinical_note = "Clinical correlation recommended."

        return AgentResult(
            agent_name=self.name,
            status="success",
            data={
                "findings": findings,
                "image_stats": regions_stats,
                "quality_score": min(0.98, 0.7 + contrast / 100),
                "clinical_note": clinical_note,
                "model_used": self.model_name,
            },
            processing_time_ms=0,
        )


class FindingInterpreterAgent(BaseAgent):
    """
    Agent 2: MedGemma Finding Interpreter
    Uses real MedGemma to interpret findings into clinical language.
    """

    def __init__(self):
        super().__init__("Finding Interpreter", "google/medgemma-4b-it")

    def process(self, input_data: Dict, context: Optional[Dict] = None) -> AgentResult:
        findings = input_data.get("findings", [])
        clinical_note = input_data.get("clinical_note", "")

        interpreted = []

        for finding in findings:
            if MODEL_LOADED:
                prompt = f"""As a radiologist, interpret this chest X-ray finding:

Finding Type: {finding.get("type")}
Region: {finding.get("region")}
Severity: {finding.get("severity")}
Description: {finding.get("description")}

Provide:
1. Clinical significance (1 sentence)
2. Top 3 differential diagnoses
3. Recommended follow-up

Be concise and clinically relevant."""

                response = generate_medgemma_response(prompt, max_tokens=200)

                interpreted.append(
                    {
                        "original": finding,
                        "medgemma_interpretation": response,
                        "clinical_significance": self._extract_significance(
                            response, finding
                        ),
                        "differential_diagnoses": self._extract_differentials(
                            response, finding
                        ),
                    }
                )
            else:
                # Demo fallback
                interpreted.append(
                    {
                        "original": finding,
                        "medgemma_interpretation": "[Model not loaded - demo mode]",
                        "clinical_significance": "Clinical correlation recommended.",
                        "differential_diagnoses": ["Requires radiologist review"],
                    }
                )

        return AgentResult(
            agent_name=self.name,
            status="success",
            data={
                "interpreted_findings": interpreted,
                "findings_count": len(findings),
                "model_used": self.model_name if MODEL_LOADED else "Demo mode",
            },
            processing_time_ms=0,
        )

    def _extract_significance(self, response: str, finding: Dict) -> str:
        # Extract first meaningful sentence from response
        sentences = response.split(".")
        if sentences:
            return sentences[0].strip() + "."
        return f"{finding.get('type', 'Finding')} requires clinical correlation."

    def _extract_differentials(self, response: str, finding: Dict) -> List[str]:
        # Default differentials based on finding type
        defaults = {
            "opacity": ["Pneumonia", "Atelectasis", "Mass/Nodule"],
            "cardiomegaly": ["Heart failure", "Cardiomyopathy", "Pericardial effusion"],
            "asymmetry": ["Pleural effusion", "Consolidation", "Mass effect"],
            "normal": ["No significant pathology"],
        }
        return defaults.get(finding.get("type", ""), ["Undetermined"])


class ReportGeneratorAgent(BaseAgent):
    """
    Agent 3: MedGemma Report Generator
    Uses real MedGemma to create structured radiology reports.
    """

    def __init__(self):
        super().__init__("Report Generator", "google/medgemma-4b-it")

    def process(self, input_data: Dict, context: Optional[Dict] = None) -> AgentResult:
        interpreted = input_data.get("interpreted_findings", [])
        clinical_history = (
            context.get("clinical_history", "Not provided")
            if context
            else "Not provided"
        )

        if MODEL_LOADED:
            # Prepare findings for MedGemma
            findings_text = ""
            for item in interpreted:
                orig = item.get("original", {})
                interp = item.get("medgemma_interpretation", "")
                findings_text += (
                    f"- {orig.get('type', 'Finding')}: {orig.get('description', '')}\n"
                )
                findings_text += f"  Interpretation: {interp[:150]}...\n"

            prompt = f"""Generate a professional radiology report for a chest X-ray with these details:

CLINICAL HISTORY: {clinical_history}

FINDINGS FROM IMAGE ANALYSIS:
{findings_text if findings_text else "No significant abnormalities detected."}

Generate a complete, structured radiology report with:
- TECHNIQUE section
- COMPARISON section  
- FINDINGS section (detailed)
- IMPRESSION section (numbered list)
- RECOMMENDATIONS

Use proper radiological terminology. Be concise but thorough."""

            report_text = generate_medgemma_response(prompt, max_tokens=500)
        else:
            report_text = self._generate_demo_report(interpreted, clinical_history)

        # Wrap in standard format
        full_report = f"""
{"=" * 80}
                         CHEST RADIOGRAPH REPORT
                    Generated by RadioFlow AI System
{"=" * 80}

CLINICAL INDICATION:
{clinical_history}

{report_text}

{"=" * 80}
⚠️ AI-GENERATED REPORT - Requires radiologist verification before clinical use.
Model: {self.model_name} | Generated: {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}
{"=" * 80}
"""

        return AgentResult(
            agent_name=self.name,
            status="success",
            data={
                "full_report": full_report.strip(),
                "findings_count": len(interpreted),
                "model_used": self.model_name if MODEL_LOADED else "Demo mode",
            },
            processing_time_ms=0,
        )

    def _generate_demo_report(
        self, interpreted: List[Dict], clinical_history: str
    ) -> str:
        findings_list = []
        for item in interpreted:
            orig = item.get("original", {})
            findings_list.append(f"- {orig.get('description', 'Finding noted')}")

        return f"""
TECHNIQUE:
Single frontal (PA) view of the chest was obtained.

COMPARISON:
None available.

FINDINGS:
LUNGS: {chr(10).join(findings_list) if findings_list else "Clear bilaterally. No focal consolidation."}

HEART: Normal cardiac silhouette size.

MEDIASTINUM: Unremarkable.

BONES: No acute osseous abnormality.

IMPRESSION:
1. {"Findings as described above require clinical correlation." if interpreted else "No acute cardiopulmonary abnormality."}

RECOMMENDATIONS:
Clinical correlation recommended as indicated.
"""


class PriorityRouterAgent(BaseAgent):
    """
    Agent 4: MedGemma Priority Router
    Uses real MedGemma to assess urgency and route cases.
    """

    PRIORITY_LEVELS = {
        "STAT": {
            "color": "#ef4444",
            "response_time": "< 30 minutes",
            "score_range": (0.8, 1.0),
        },
        "URGENT": {
            "color": "#f59e0b",
            "response_time": "< 4 hours",
            "score_range": (0.5, 0.8),
        },
        "ROUTINE": {
            "color": "#22c55e",
            "response_time": "< 24 hours",
            "score_range": (0.0, 0.5),
        },
    }

    def __init__(self):
        super().__init__("Priority Router", "google/medgemma-4b-it")

    def process(self, input_data: Dict, context: Optional[Dict] = None) -> AgentResult:
        full_report = input_data.get("full_report", "")
        original_findings = context.get("original_findings", []) if context else []

        # Calculate base priority score from findings
        severity_scores = {
            "critical": 1.0,
            "high": 0.8,
            "moderate": 0.5,
            "mild": 0.3,
            "none": 0.1,
        }
        max_severity = 0.2
        for finding in original_findings:
            sev = finding.get("severity", "none")
            max_severity = max(max_severity, severity_scores.get(sev, 0.2))

        if MODEL_LOADED:
            # Use MedGemma for clinical priority assessment
            prompt = f"""As a radiologist, assess the clinical priority of this chest X-ray report:

{full_report[:1000]}

Based on the findings, determine:
1. PRIORITY LEVEL: STAT (immediate), URGENT (within 4 hours), or ROUTINE (within 24 hours)
2. CRITICAL FINDINGS: List any findings requiring immediate physician notification
3. RECOMMENDED ACTIONS: What should happen next?

Respond concisely."""

            medgemma_assessment = generate_medgemma_response(prompt, max_tokens=200)

            # Adjust score based on MedGemma's assessment
            if (
                "STAT" in medgemma_assessment.upper()
                or "IMMEDIATE" in medgemma_assessment.upper()
            ):
                max_severity = max(max_severity, 0.85)
            elif "URGENT" in medgemma_assessment.upper():
                max_severity = max(max_severity, 0.55)
        else:
            medgemma_assessment = "Priority assessment based on finding severity."

        # Determine priority level
        priority_level = "ROUTINE"
        if max_severity >= 0.8:
            priority_level = "STAT"
        elif max_severity >= 0.5:
            priority_level = "URGENT"

        return AgentResult(
            agent_name=self.name,
            status="success",
            data={
                "priority_score": round(max_severity, 2),
                "priority_level": priority_level,
                "priority_details": self.PRIORITY_LEVELS[priority_level],
                "medgemma_assessment": medgemma_assessment,
                "routing_recommendation": {
                    "destination": f"{priority_level} Reading Queue",
                    "notification_required": priority_level in ["STAT", "URGENT"],
                },
                "model_used": self.model_name if MODEL_LOADED else "Demo mode",
            },
            processing_time_ms=0,
        )


print("βœ… All agents defined with real MedGemma integration")

# %% [markdown]
# ## 6. Workflow Orchestrator


# %%
@dataclass
class WorkflowResult:
    """Complete result from RadioFlow workflow"""

    workflow_id: str
    status: str
    total_duration_ms: float
    final_report: str = ""
    priority_level: str = "ROUTINE"
    priority_score: float = 0.0
    findings_count: int = 0
    agent_results: Dict[str, AgentResult] = field(default_factory=dict)


class RadioFlowOrchestrator:
    """Main orchestrator for the RadioFlow multi-agent system."""

    def __init__(self):
        self.agents = {
            "cxr_analyzer": CXRAnalyzerAgent(),
            "finding_interpreter": FindingInterpreterAgent(),
            "report_generator": ReportGeneratorAgent(),
            "priority_router": PriorityRouterAgent(),
        }
        print("πŸš€ RadioFlow Orchestrator initialized with 4 agents")

    def process(
        self, image: Image.Image, context: Optional[Dict] = None
    ) -> WorkflowResult:
        start = time.time()
        workflow_id = f"rf_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
        context = context or {}

        print(f"\n{'=' * 60}")
        print(f"🩻 RadioFlow Workflow: {workflow_id}")
        print(f"   Model: {'MedGemma (REAL)' if MODEL_LOADED else 'Demo Mode'}")
        print(f"{'=' * 60}")

        # Stage 1: CXR Analysis
        print("\nπŸ”¬ Stage 1: Analyzing chest X-ray...")
        cxr_result = self.agents["cxr_analyzer"](image, context)
        findings = cxr_result.data.get("findings", [])
        print(f"   βœ… Detected {len(findings)} finding(s)")
        for f in findings[:3]:
            print(f"      β€’ {f['type']}: {f['description'][:50]}...")

        # Stage 2: Finding Interpretation
        print("\nπŸ“‹ Stage 2: Interpreting findings with MedGemma...")
        interp_result = self.agents["finding_interpreter"](cxr_result.data, context)
        print(f"   βœ… Generated clinical interpretations")

        # Stage 3: Report Generation
        print("\nπŸ“ Stage 3: Generating radiology report...")
        report_result = self.agents["report_generator"](interp_result.data, context)
        print(
            f"   βœ… Report generated ({len(report_result.data.get('full_report', ''))} chars)"
        )

        # Stage 4: Priority Routing
        print("\n🚦 Stage 4: Assessing priority...")
        priority_context = {**context, "original_findings": findings}
        priority_result = self.agents["priority_router"](
            report_result.data, priority_context
        )
        level = priority_result.data.get("priority_level")
        score = priority_result.data.get("priority_score", 0)
        print(f"   βœ… Priority: {level} ({score:.0%})")

        total_time = (time.time() - start) * 1000

        print(f"\n{'=' * 60}")
        print(f"βœ… Workflow Complete in {total_time:.0f}ms")
        print(f"{'=' * 60}\n")

        return WorkflowResult(
            workflow_id=workflow_id,
            status="success",
            total_duration_ms=total_time,
            final_report=report_result.data.get("full_report", ""),
            priority_level=level,
            priority_score=score,
            findings_count=len(findings),
            agent_results={
                "cxr_analyzer": cxr_result,
                "finding_interpreter": interp_result,
                "report_generator": report_result,
                "priority_router": priority_result,
            },
        )


orchestrator = RadioFlowOrchestrator()

# %% [markdown]
# ## 7. Run Demo with Your Own Image
#
# ### Option A: Upload your own X-ray image
# 1. Click "Add data" in the right panel β†’ "Upload" β†’ Select your X-ray image
# 2. Set `USE_CUSTOM_IMAGE = True` below
# 3. Update `CUSTOM_IMAGE_PATH` with your image filename
#
# ### Option B: Use generated sample image
# Keep `USE_CUSTOM_IMAGE = False` to use the auto-generated sample


# %%
# ========== CONFIGURATION - EDIT THIS ==========
USE_CUSTOM_IMAGE = False  # Set to True to use your own image
CUSTOM_IMAGE_PATH = "/kaggle/input/your-dataset/your-xray.jpg"  # Update this path
# ===============================================


def create_sample_cxr(size=(512, 512), seed=None):
    """Create a simulated chest X-ray for demo purposes."""
    if seed:
        np.random.seed(seed)

    img = Image.new("L", size, color=30)
    draw = ImageDraw.Draw(img)

    w, h = size

    # Lung fields (darker areas)
    draw.ellipse([50, 80, w // 2 - 20, h - 50], fill=20)  # Left lung
    draw.ellipse(
        [w // 2 + 20, 80, w - 50, h - 50], fill=25
    )  # Right lung (slightly denser)

    # Heart shadow (bright/dense)
    draw.ellipse([w // 3, h // 3, 2 * w // 3, 2 * h // 3], fill=80)

    # Spine
    draw.rectangle([w // 2 - 15, 50, w // 2 + 15, h - 30], fill=90)

    # Ribs
    for i in range(8):
        y = 100 + i * 45
        draw.arc([30, y, w // 2 - 30, y + 40], 180, 360, fill=70, width=2)
        draw.arc([w // 2 + 30, y, w - 30, y + 40], 180, 360, fill=70, width=2)

    # Add some noise
    img_array = np.array(img)
    noise = np.random.normal(0, 5, img_array.shape)
    img_array = np.clip(img_array + noise, 0, 255).astype(np.uint8)

    return Image.fromarray(img_array).convert("RGB")


# Load image based on configuration
if USE_CUSTOM_IMAGE:
    print(f"πŸ“‚ Loading custom image: {CUSTOM_IMAGE_PATH}")
    try:
        sample_image = Image.open(CUSTOM_IMAGE_PATH).convert("RGB")
        # Resize if too large
        max_size = 1024
        if max(sample_image.size) > max_size:
            sample_image.thumbnail((max_size, max_size), Image.LANCZOS)
        print(f"   βœ… Image loaded! Size: {sample_image.size}")
        title = "Your Chest X-Ray"
    except Exception as e:
        print(f"   ❌ Error loading image: {e}")
        print("   Falling back to sample image...")
        sample_image = create_sample_cxr(seed=42)
        title = "Sample Chest X-Ray (fallback)"
else:
    print("🎨 Using generated sample image")
    sample_image = create_sample_cxr(seed=42)
    title = "Generated Sample Chest X-Ray"

# Display
plt.figure(figsize=(8, 8))
plt.imshow(sample_image, cmap="gray")
plt.title(title, fontsize=14)
plt.axis("off")
plt.show()

# %%
# Run the workflow
clinical_context = {
    "clinical_history": "65-year-old male presenting with productive cough and low-grade fever for 5 days. History of hypertension and type 2 diabetes.",
    "symptoms": "Cough, fever, mild dyspnea on exertion",
}

print("🩻 Processing chest X-ray with RadioFlow...\n")
result = orchestrator.process(sample_image, clinical_context)

# %% [markdown]
# ## 8. Results

# %%
# Display the generated report
print(result.final_report)

# %%
# Priority Assessment Display
priority_data = result.agent_results["priority_router"].data
colors = {"STAT": "#ef4444", "URGENT": "#f59e0b", "ROUTINE": "#22c55e"}

display(
    HTML(f"""
<div style="padding: 25px; background: linear-gradient(135deg, #1e3a5f, #2d5a87); 
            border-radius: 15px; color: white; margin: 20px 0; box-shadow: 0 4px 6px rgba(0,0,0,0.3);">
    <h2 style="margin: 0 0 20px 0; font-size: 24px;">🚦 Priority Assessment</h2>
    <div style="display: flex; gap: 40px; flex-wrap: wrap;">
        <div style="text-align: center;">
            <div style="font-size: 48px; font-weight: bold; color: {colors.get(result.priority_level, "#fff")};">
                {result.priority_level}
            </div>
            <div style="opacity: 0.8; font-size: 14px;">Priority Level</div>
        </div>
        <div style="text-align: center;">
            <div style="font-size: 48px; font-weight: bold;">{result.priority_score:.0%}</div>
            <div style="opacity: 0.8; font-size: 14px;">Urgency Score</div>
        </div>
        <div style="text-align: center;">
            <div style="font-size: 48px; font-weight: bold;">{result.findings_count}</div>
            <div style="opacity: 0.8; font-size: 14px;">Findings</div>
        </div>
        <div style="text-align: center;">
            <div style="font-size: 48px; font-weight: bold;">{result.total_duration_ms / 1000:.1f}s</div>
            <div style="opacity: 0.8; font-size: 14px;">Total Time</div>
        </div>
    </div>
    <div style="margin-top: 20px; padding: 15px; background: rgba(255,255,255,0.1); border-radius: 8px;">
        <strong>MedGemma Assessment:</strong><br>
        {priority_data.get("medgemma_assessment", "N/A")[:300]}
    </div>
</div>
""")
)

# %%
# Agent Metrics
metrics_data = []
for key, agent_result in result.agent_results.items():
    metrics_data.append(
        {
            "Agent": agent_result.agent_name,
            "Status": "βœ… Success" if agent_result.status == "success" else "❌ Error",
            "Time (ms)": f"{agent_result.processing_time_ms:.0f}",
            "Model": agent_result.data.get("model_used", "N/A")[:40],
        }
    )

metrics_df = pd.DataFrame(metrics_data)
print("\nπŸ“Š Agent Performance Metrics:")
display(metrics_df)

# %%
# Create workflow visualization
fig = go.Figure()

agents = ["CXR Analyzer", "Finding Interpreter", "Report Generator", "Priority Router"]
times = [
    result.agent_results["cxr_analyzer"].processing_time_ms,
    result.agent_results["finding_interpreter"].processing_time_ms,
    result.agent_results["report_generator"].processing_time_ms,
    result.agent_results["priority_router"].processing_time_ms,
]

fig.add_trace(
    go.Bar(
        x=times,
        y=agents,
        orientation="h",
        marker_color=["#3b82f6", "#8b5cf6", "#10b981", "#f59e0b"],
        text=[f"{t:.0f}ms" for t in times],
        textposition="inside",
        textfont=dict(color="white", size=14),
    )
)

fig.update_layout(
    title="Agent Processing Times",
    xaxis_title="Time (ms)",
    height=300,
    margin=dict(l=150, r=40, t=60, b=40),
    paper_bgcolor="rgba(0,0,0,0)",
    plot_bgcolor="rgba(0,0,0,0)",
)

fig.show()

# %% [markdown]
# ## 9. MedGemma Interpretation Showcase

# %%
# Show MedGemma's clinical interpretations
print("🧠 MedGemma Clinical Interpretations:\n")
print("=" * 60)

interpreted = result.agent_results["finding_interpreter"].data.get(
    "interpreted_findings", []
)
for i, item in enumerate(interpreted, 1):
    orig = item.get("original", {})
    interp = item.get("medgemma_interpretation", "")

    print(f"\nπŸ“‹ Finding {i}: {orig.get('type', 'Unknown').upper()}")
    print(f"   Region: {orig.get('region', 'N/A')}")
    print(f"   Severity: {orig.get('severity', 'N/A')}")
    print(f"\n   πŸ€– MedGemma Interpretation:")
    print(f"   {interp[:500]}")
    print("-" * 60)

# %% [markdown]
# ## 10. Conclusion
#
# ### βœ… Key Technical Achievements
#
# 1. **Real MedGemma Integration**: This notebook uses the actual MedGemma-4B model for clinical
#    interpretation, report generation, and priority assessment - not simulated responses.
#
# 2. **Multi-Agent Architecture**: Successfully implemented a 4-agent pipeline demonstrating
#    agentic workflow principles with clear separation of concerns.
#
# 3. **Efficient Inference**: Uses 4-bit quantization (bitsandbytes) to run MedGemma on
#    Kaggle's free T4 GPU within memory constraints.
#
# 4. **Production-Ready**: Generates professional radiology reports following clinical standards.
#
# ### πŸ“Š Competition Alignment
#
# | Criterion | How RadioFlow Addresses It |
# |-----------|---------------------------|
# | **Effective HAI-DEF Use** | Real MedGemma inference throughout pipeline |
# | **Problem Domain** | Addresses radiologist burnout and workflow inefficiency |
# | **Impact Potential** | Quantifiable time savings and improved critical finding detection |
# | **Product Feasibility** | Deployable demo with clear technical architecture |
# | **Agentic Workflow** | 4-agent orchestrated system with handoffs |
#
# ---
#
# **πŸ”— Live Demo:** https://huggingface.co/spaces/SamarpeetGarad/radioflow
#
# **Thank you for reviewing the RadioFlow submission!** πŸ™

# %%
print("\n" + "=" * 60)
print("πŸ† RadioFlow - MedGemma Impact Challenge Submission")
print("=" * 60)
print(f"\nπŸ“Š Final Summary:")
print(f"   β€’ Model Used: {'MedGemma-4B (REAL)' if MODEL_LOADED else 'Demo Mode'}")
print(f"   β€’ Total Processing Time: {result.total_duration_ms:.0f}ms")
print(f"   β€’ Findings Detected: {result.findings_count}")
print(f"   β€’ Priority Level: {result.priority_level}")
print(f"   β€’ Priority Score: {result.priority_score:.0%}")
print(f"\nπŸ”— Live Demo: https://huggingface.co/spaces/SamarpeetGarad/radioflow")
print("\nβœ… Notebook Complete!")