File size: 36,700 Bytes
86f402d c376e14 86f402d c376e14 86f402d 1a97904 da343a7 86f402d da343a7 86f402d da343a7 86f402d da343a7 1a97904 86f402d bb7e939 86f402d 58a4476 86f402d b08f876 bb7e939 b08f876 bb7e939 86f402d 0989643 86f402d 58a4476 86f402d b08f876 58a4476 86f402d 1a97904 58a4476 c376e14 58a4476 86f402d c376e14 86f402d 4af4003 86f402d 4af4003 86f402d 5241b71 86f402d 4af4003 86f402d 4af4003 86f402d 4af4003 86f402d 4af4003 86f402d | 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 | """
MedGemma Agent - LLM agent with tool calling and staged thinking feedback
Pipeline: MedGemma independent exam → Tools (MONET/ConvNeXt/GradCAM) → MedGemma reconciliation → Management
"""
import sys
import time
import random
import json
import os
import subprocess
import threading
from typing import Optional, Generator, Dict, Any
from PIL import Image
class MCPClient:
"""
Minimal MCP client that communicates with a FastMCP subprocess over stdio.
Uses raw newline-delimited JSON-RPC 2.0 so the main process (Python 3.9)
does not need the mcp library. The subprocess is launched with python3.11
which has mcp installed.
"""
def __init__(self):
self._process = None
self._lock = threading.Lock()
self._id_counter = 0
def _next_id(self) -> int:
self._id_counter += 1
return self._id_counter
def _send(self, obj: dict):
line = json.dumps(obj) + "\n"
self._process.stdin.write(line)
self._process.stdin.flush()
def _recv(self, timeout: int = 300) -> dict:
import select
deadline = time.time() + timeout
while True:
remaining = deadline - time.time()
if remaining <= 0:
raise RuntimeError(
f"MCP server did not respond within {timeout}s"
)
ready, _, _ = select.select(
[self._process.stdout], [], [], min(remaining, 5)
)
if not ready:
# Check if subprocess died
if self._process.poll() is not None:
raise RuntimeError(
f"MCP server exited with code {self._process.returncode}"
)
continue
line = self._process.stdout.readline()
if not line:
raise RuntimeError("MCP server closed connection unexpectedly")
line = line.strip()
if not line:
continue
msg = json.loads(line)
# Skip server-initiated notifications (no "id" key)
if "id" in msg:
return msg
def _initialize(self):
"""Send MCP initialize handshake."""
req_id = self._next_id()
self._send({
"jsonrpc": "2.0",
"id": req_id,
"method": "initialize",
"params": {
"protocolVersion": "2024-11-05",
"capabilities": {},
"clientInfo": {"name": "SkinProAI", "version": "1.0.0"},
},
})
self._recv() # consume initialize response
# Confirm initialization
self._send({
"jsonrpc": "2.0",
"method": "notifications/initialized",
"params": {},
})
def start(self):
"""Spawn the MCP server subprocess and complete the handshake."""
root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
server_script = os.path.join(root, "mcp_server", "server.py")
# Force MCP tool models (MONET, ConvNeXt) onto CPU so they don't
# compete with MedGemma for GPU VRAM (T4 has only 16 GB).
env = os.environ.copy()
env["SKINPRO_TOOL_DEVICE"] = "cpu"
# stderr inherits parent's stderr (not PIPE) to avoid deadlock:
# when the subprocess downloads models, progress output fills the
# 64KB pipe buffer and blocks the subprocess while the main process
# is blocked waiting on stdout — classic subprocess deadlock.
self._process = subprocess.Popen(
[sys.executable, server_script],
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
stderr=None, # inherit parent stderr
text=True,
bufsize=1,
env=env,
)
self._initialize()
def call_tool_sync(self, tool_name: str, arguments: dict) -> dict:
"""Call a tool synchronously and return the parsed result dict."""
with self._lock:
req_id = self._next_id()
self._send({
"jsonrpc": "2.0",
"id": req_id,
"method": "tools/call",
"params": {"name": tool_name, "arguments": arguments},
})
response = self._recv()
# Protocol-level error (e.g. unknown method)
if "error" in response:
raise RuntimeError(
f"MCP tool '{tool_name}' failed: {response['error']}"
)
result = response["result"]
content_text = result["content"][0]["text"]
# Tool-level error (isError=True means the tool itself raised an exception)
if result.get("isError"):
raise RuntimeError(f"MCP tool '{tool_name}' error: {content_text}")
return json.loads(content_text)
def stop(self):
"""Terminate the MCP server subprocess."""
if self._process:
try:
self._process.stdin.close()
self._process.terminate()
self._process.wait(timeout=5)
except Exception:
pass
self._process = None
# Rotating verbs for spinner effect
ANALYSIS_VERBS = [
"Analyzing", "Examining", "Processing", "Inspecting", "Evaluating",
"Scanning", "Assessing", "Reviewing", "Studying", "Interpreting"
]
# Comprehensive visual exam prompt (combined from 4 separate stages)
COMPREHENSIVE_EXAM_PROMPT = """Perform a systematic dermoscopic examination of this skin lesion. Assess ALL of the following in a SINGLE concise analysis:
1. PATTERN: Overall architecture, symmetry (symmetric/asymmetric), organization
2. COLORS: List all colors present (brown, black, blue, white, red, pink) and distribution
3. BORDER: Sharp vs gradual, regular vs irregular, any disruptions
4. STRUCTURES: Pigment network, dots/globules, streaks, blue-white veil, regression, vessels
Then provide:
- Top 3 differential diagnoses with brief reasoning
- Concern level (1-5, where 5=urgent)
- Single most important feature driving your assessment
Be CONCISE - focus on clinically relevant findings only."""
def get_verb():
"""Get a random analysis verb for spinner effect"""
return random.choice(ANALYSIS_VERBS)
class MedGemmaAgent:
"""
Medical image analysis agent with:
- Staged thinking display (no emojis)
- Tool calling (MONET, ConvNeXt, Grad-CAM)
- Streaming responses
"""
def __init__(self, verbose: bool = True):
self.verbose = verbose
self.pipe = None
self.model_id = "google/medgemma-4b-it"
self.loaded = False
# Tools (legacy direct instances, kept for fallback / non-MCP use)
self.monet_tool = None
self.convnext_tool = None
self.gradcam_tool = None
self.rag_tool = None
self.tools_loaded = False
# MCP client
self.mcp_client = None
# State for confirmation flow
self.last_diagnosis = None
self.last_monet_result = None
self.last_image = None
self.last_medgemma_exam = None # Store independent MedGemma findings
self.last_reconciliation = None
def reset_state(self):
"""Reset analysis state for new analysis (keeps models loaded)"""
self.last_diagnosis = None
self.last_monet_result = None
self.last_image = None
self.last_medgemma_exam = None
self.last_reconciliation = None
def _print(self, message: str):
"""Print if verbose"""
if self.verbose:
print(message, flush=True)
def load_model(self):
"""Load MedGemma model"""
if self.loaded:
return
self._print("Initializing MedGemma agent...")
import os
import torch
from transformers import pipeline, AutoProcessor
# Authenticate with HF Hub (required for gated models like MedGemma)
hf_token = os.environ.get("HF_TOKEN")
if hf_token:
from huggingface_hub import login
login(token=hf_token, add_to_git_credential=False)
self._print("Authenticated with HF Hub")
else:
self._print("Warning: HF_TOKEN not set — gated models will fail")
self._print(f"Loading model: {self.model_id}")
if torch.cuda.is_available():
device = "cuda"
self._print(f"Using GPU: {torch.cuda.get_device_name(0)}")
elif torch.backends.mps.is_available():
device = "mps"
self._print("Using Apple Silicon (MPS)")
else:
device = "cpu"
self._print("Using CPU")
model_kwargs = dict(
dtype=torch.bfloat16, # bfloat16 on all devices (halves RAM on CPU: ~8GB vs ~16GB)
device_map="auto",
)
start = time.time()
processor = AutoProcessor.from_pretrained(
self.model_id, use_fast=True, token=hf_token,
)
self.pipe = pipeline(
"image-text-to-text",
model=self.model_id,
model_kwargs=model_kwargs,
token=hf_token, # pass explicitly in addition to login()
image_processor=processor.image_processor,
tokenizer=processor.tokenizer,
)
# Clear default max_length (20) from generation_config to avoid conflict
# with max_new_tokens passed at inference time
if hasattr(self.pipe.model, "generation_config"):
self.pipe.model.generation_config.max_length = None
self._print(f"Model loaded in {time.time() - start:.1f}s")
self.loaded = True
def load_tools(self):
"""Load tool models (MONET + ConvNeXt + Grad-CAM + RAG)"""
if self.tools_loaded:
return
from models.monet_tool import MonetTool
self.monet_tool = MonetTool()
self.monet_tool.load()
from models.convnext_classifier import ConvNeXtClassifier
self.convnext_tool = ConvNeXtClassifier()
self.convnext_tool.load()
from models.gradcam_tool import GradCAMTool
self.gradcam_tool = GradCAMTool(classifier=self.convnext_tool)
self.gradcam_tool.load()
from models.guidelines_rag import get_guidelines_rag
self.rag_tool = get_guidelines_rag()
if not self.rag_tool.loaded:
self.rag_tool.load_index()
self.tools_loaded = True
def load_tools_via_mcp(self):
"""Start the MCP server subprocess and mark tools as loaded."""
if self.tools_loaded:
return
self.mcp_client = MCPClient()
self.mcp_client.start()
self._print("MCP server started successfully")
self.tools_loaded = True
def _multi_pass_visual_exam(self, image, question: Optional[str] = None) -> Generator[str, None, Dict[str, str]]:
"""
MedGemma performs comprehensive visual examination BEFORE tools run.
Single prompt covers pattern, colors, borders, structures, and differentials.
Returns findings dict after yielding all output.
"""
findings = {}
yield f"\n[STAGE:medgemma_exam]MedGemma Visual Examination[/STAGE]\n"
yield f"[THINKING]Performing systematic dermoscopic assessment...[/THINKING]\n"
# Build prompt with optional clinical question
exam_prompt = COMPREHENSIVE_EXAM_PROMPT
if question:
exam_prompt += f"\n\nCLINICAL QUESTION: {question}"
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": image},
{"type": "text", "text": exam_prompt}
]
}
]
try:
time.sleep(0.2)
output = self.pipe(messages, max_new_tokens=200)
result = output[0]["generated_text"][-1]["content"]
findings['synthesis'] = result
yield f"[RESPONSE]\n"
words = result.split()
for i, word in enumerate(words):
time.sleep(0.015)
yield word + (" " if i < len(words) - 1 else "")
yield f"\n[/RESPONSE]\n"
except Exception as e:
findings['synthesis'] = f"Analysis failed: {e}"
yield f"[ERROR]Visual examination failed: {e}[/ERROR]\n"
self.last_medgemma_exam = findings
return findings
def _reconcile_findings(
self,
image,
medgemma_exam: Dict[str, str],
monet_result: Dict[str, Any],
convnext_result: Dict[str, Any],
question: Optional[str] = None
) -> Generator[str, None, None]:
"""
MedGemma reconciles its independent findings with tool outputs.
Identifies agreements, disagreements, and provides integrated assessment.
"""
yield f"\n[STAGE:reconciliation]Reconciling MedGemma Findings with Tool Results[/STAGE]\n"
yield f"[THINKING]Comparing independent visual assessment against AI classification tools...[/THINKING]\n"
top = convnext_result['predictions'][0]
runner_up = convnext_result['predictions'][1] if len(convnext_result['predictions']) > 1 else None
# Build MONET features string
monet_top = sorted(monet_result["features"].items(), key=lambda x: x[1], reverse=True)[:5]
monet_str = ", ".join([f"{k.replace('MONET_', '').replace('_', ' ')}: {v:.0%}" for k, v in monet_top])
reconciliation_prompt = f"""You performed an independent visual examination of this lesion and concluded:
YOUR ASSESSMENT:
{medgemma_exam.get('synthesis', 'Not available')[:600]}
The AI classification tools produced these results:
- ConvNeXt classifier: {top['full_name']} ({top['probability']:.1%} confidence)
{f"- Runner-up: {runner_up['full_name']} ({runner_up['probability']:.1%})" if runner_up else ""}
- Key MONET features: {monet_str}
{f'CLINICAL QUESTION: {question}' if question else ''}
Reconcile your visual findings with the AI classification:
1. AGREEMENT/DISAGREEMENT: Do your findings support the AI diagnosis? Any conflicts?
2. INTEGRATED ASSESSMENT: Final diagnosis considering all evidence
3. CONFIDENCE (1-10): How certain? What would change your assessment?
Be concise and specific."""
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": image},
{"type": "text", "text": reconciliation_prompt}
]
}
]
try:
output = self.pipe(messages, max_new_tokens=150)
reconciliation = output[0]["generated_text"][-1]["content"]
self.last_reconciliation = reconciliation
yield f"[RESPONSE]\n"
words = reconciliation.split()
for i, word in enumerate(words):
time.sleep(0.015)
yield word + (" " if i < len(words) - 1 else "")
yield f"\n[/RESPONSE]\n"
except Exception as e:
yield f"[ERROR]Reconciliation failed: {e}[/ERROR]\n"
def analyze_image_stream(
self,
image_path: str,
question: Optional[str] = None,
max_tokens: int = 512,
use_tools: bool = True
) -> Generator[str, None, None]:
"""
Stream analysis with new pipeline:
1. MedGemma independent multi-pass exam
2. MONET + ConvNeXt + GradCAM tools
3. MedGemma reconciliation
4. Confirmation request
"""
if not self.loaded:
yield "[STAGE:loading]Initializing MedGemma...[/STAGE]\n"
self.load_model()
yield f"[STAGE:image]{get_verb()} image...[/STAGE]\n"
try:
image = Image.open(image_path).convert("RGB")
self.last_image = image
except Exception as e:
yield f"[ERROR]Failed to load image: {e}[/ERROR]\n"
return
# Load tools early via MCP subprocess
if use_tools and not self.tools_loaded:
yield f"[STAGE:tools]Loading analysis tools...[/STAGE]\n"
self.load_tools_via_mcp()
# ===== PHASE 1: MedGemma Independent Visual Examination =====
medgemma_exam = {}
for chunk in self._multi_pass_visual_exam(image, question):
yield chunk
if isinstance(chunk, dict):
medgemma_exam = chunk
medgemma_exam = self.last_medgemma_exam or {}
monet_result = None
convnext_result = None
if use_tools:
# ===== PHASE 2: Run Classification Tools =====
yield f"\n[STAGE:tools_run]Running AI Classification Tools[/STAGE]\n"
yield f"[THINKING]Now running MONET and ConvNeXt to compare against visual examination...[/THINKING]\n"
# MONET Feature Extraction
time.sleep(0.2)
yield f"\n[STAGE:monet]MONET Feature Extraction[/STAGE]\n"
try:
monet_result = self.mcp_client.call_tool_sync(
"monet_analyze", {"image_path": image_path}
)
self.last_monet_result = monet_result
yield f"[TOOL_OUTPUT:MONET Features]\n"
for name, score in monet_result["features"].items():
short_name = name.replace("MONET_", "").replace("_", " ").title()
bar_filled = int(score * 10)
bar = "|" + "=" * bar_filled + "-" * (10 - bar_filled) + "|"
yield f" {short_name}: {bar} {score:.0%}\n"
yield f"[/TOOL_OUTPUT]\n"
except Exception as e:
yield f"[ERROR]MONET failed: {e}[/ERROR]\n"
# ConvNeXt Classification
time.sleep(0.2)
yield f"\n[STAGE:convnext]ConvNeXt Classification[/STAGE]\n"
try:
monet_scores = monet_result["vector"] if monet_result else None
convnext_result = self.mcp_client.call_tool_sync(
"classify_lesion",
{
"image_path": image_path,
"monet_scores": monet_scores,
},
)
self.last_diagnosis = convnext_result
yield f"[TOOL_OUTPUT:Classification Results]\n"
for pred in convnext_result["predictions"][:5]:
prob = pred['probability']
bar_filled = int(prob * 20)
bar = "|" + "=" * bar_filled + "-" * (20 - bar_filled) + "|"
yield f" {pred['class']}: {bar} {prob:.1%}\n"
yield f" {pred['full_name']}\n"
yield f"[/TOOL_OUTPUT]\n"
top = convnext_result['predictions'][0]
yield f"[RESULT]ConvNeXt Primary: {top['full_name']} ({top['probability']:.1%})[/RESULT]\n"
except Exception as e:
yield f"[ERROR]ConvNeXt failed: {e}[/ERROR]\n"
# Grad-CAM Visualization
time.sleep(0.2)
yield f"\n[STAGE:gradcam]Grad-CAM Attention Map[/STAGE]\n"
try:
gradcam_result = self.mcp_client.call_tool_sync(
"generate_gradcam", {"image_path": image_path}
)
gradcam_path = gradcam_result["gradcam_path"]
yield f"[GRADCAM_IMAGE:{gradcam_path}]\n"
except Exception as e:
yield f"[ERROR]Grad-CAM failed: {e}[/ERROR]\n"
# ===== PHASE 3: MedGemma Reconciliation =====
if convnext_result and monet_result and medgemma_exam:
for chunk in self._reconcile_findings(
image, medgemma_exam, monet_result, convnext_result, question
):
yield chunk
# ===== PHASE 4: Auto Management Guidance (with RAG) =====
if convnext_result and self.mcp_client:
for chunk in self.generate_management_guidance():
yield chunk
def generate_management_guidance(
self,
user_confirmed: bool = True,
user_feedback: Optional[str] = None
) -> Generator[str, None, None]:
"""
Generate LESION-SPECIFIC management guidance using RAG + MedGemma reasoning.
References specific findings from this analysis, not generic textbook management.
"""
if not self.last_diagnosis:
yield "[ERROR]No diagnosis available. Please analyze an image first.[/ERROR]\n"
return
top = self.last_diagnosis['predictions'][0]
runner_up = self.last_diagnosis['predictions'][1] if len(self.last_diagnosis['predictions']) > 1 else None
diagnosis = top['full_name']
if not user_confirmed and user_feedback:
yield f"[THINKING]Clinician provided alternative assessment: {user_feedback}[/THINKING]\n"
diagnosis = user_feedback
# Stage: RAG Search
time.sleep(0.3)
yield f"\n[STAGE:guidelines]Searching clinical guidelines for {diagnosis}...[/STAGE]\n"
# Get RAG context via MCP
features_desc = self.last_monet_result.get('description', '') if self.last_monet_result else ''
rag_data = self.mcp_client.call_tool_sync(
"search_guidelines",
{"query": features_desc, "diagnosis": diagnosis},
)
context = rag_data["context"]
references = rag_data["references"]
# Check guideline relevance
has_relevant_guidelines = False
if references:
diagnosis_lower = diagnosis.lower()
for ref in references:
source_lower = ref['source'].lower()
if any(term in diagnosis_lower for term in ['melanoma']) and 'melanoma' in source_lower:
has_relevant_guidelines = True
break
elif 'actinic' in diagnosis_lower and 'actinic' in source_lower:
has_relevant_guidelines = True
break
elif ref.get('score', 0) > 0.7:
has_relevant_guidelines = True
break
if not references or not has_relevant_guidelines:
yield f"[THINKING]No specific published guidelines for {diagnosis}. Using clinical knowledge.[/THINKING]\n"
context = "No specific clinical guidelines available."
references = []
# Build MONET features for context
monet_features = ""
if self.last_monet_result:
top_features = sorted(self.last_monet_result["features"].items(), key=lambda x: x[1], reverse=True)[:5]
monet_features = ", ".join([f"{k.replace('MONET_', '').replace('_', ' ')}: {v:.0%}" for k, v in top_features])
# Stage: Lesion-Specific Management Reasoning
time.sleep(0.3)
yield f"\n[STAGE:management]Generating Lesion-Specific Management Plan[/STAGE]\n"
yield f"[THINKING]Creating management plan tailored to THIS lesion's specific characteristics...[/THINKING]\n"
management_prompt = f"""Generate a CONCISE management plan for this lesion:
DIAGNOSIS: {diagnosis} ({top['probability']:.1%})
{f"Alternative: {runner_up['full_name']} ({runner_up['probability']:.1%})" if runner_up else ""}
KEY FEATURES: {monet_features}
{f"GUIDELINES: {context[:800]}" if context else ""}
Provide:
1. RECOMMENDED ACTION: Biopsy, excision, monitoring, or discharge - with specific reasoning
2. URGENCY: Routine vs urgent vs same-day referral
3. KEY CONCERNS: What features drive this recommendation
Be specific to THIS lesion. 3-5 sentences maximum."""
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": self.last_image},
{"type": "text", "text": management_prompt}
]
}
]
# Generate response
start = time.time()
try:
output = self.pipe(messages, max_new_tokens=150)
response = output[0]["generated_text"][-1]["content"]
yield f"[RESPONSE]\n"
words = response.split()
for i, word in enumerate(words):
time.sleep(0.015)
yield word + (" " if i < len(words) - 1 else "")
yield f"\n[/RESPONSE]\n"
except Exception as e:
yield f"[ERROR]Management generation failed: {e}[/ERROR]\n"
# Output references (pre-formatted by MCP server)
if references:
yield rag_data["references_display"]
yield f"\n[COMPLETE]Lesion-specific management plan generated in {time.time() - start:.1f}s[/COMPLETE]\n"
# Store response for recommendation extraction
self.last_management_response = response
def extract_recommendation(self) -> Generator[str, None, Dict[str, Any]]:
"""
Extract structured recommendation from management guidance.
Determines: BIOPSY, EXCISION, FOLLOWUP, or DISCHARGE
For BIOPSY/EXCISION, gets coordinates from MedGemma.
"""
if not self.last_management_response or not self.last_image:
yield "[ERROR]No management guidance available[/ERROR]\n"
return {"action": "UNKNOWN"}
yield f"\n[STAGE:recommendation]Extracting Clinical Recommendation[/STAGE]\n"
# Ask MedGemma to classify the recommendation
classification_prompt = f"""Based on the management plan you just provided:
{self.last_management_response[:1000]}
Classify the PRIMARY recommended action into exactly ONE of these categories:
- BIOPSY: If punch biopsy, shave biopsy, or incisional biopsy is recommended
- EXCISION: If complete surgical excision is recommended
- FOLLOWUP: If monitoring with repeat photography/dermoscopy is recommended
- DISCHARGE: If the lesion is clearly benign and no follow-up needed
Respond with ONLY the category name (BIOPSY, EXCISION, FOLLOWUP, or DISCHARGE) on the first line.
Then on the second line, provide a brief (1 sentence) justification."""
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": self.last_image},
{"type": "text", "text": classification_prompt}
]
}
]
try:
output = self.pipe(messages, max_new_tokens=100)
response = output[0]["generated_text"][-1]["content"].strip()
lines = response.split('\n')
action = lines[0].strip().upper()
justification = lines[1].strip() if len(lines) > 1 else ""
# Validate action
valid_actions = ["BIOPSY", "EXCISION", "FOLLOWUP", "DISCHARGE"]
if action not in valid_actions:
# Try to extract from response
for valid in valid_actions:
if valid in response.upper():
action = valid
break
else:
action = "FOLLOWUP" # Default to safe option
yield f"[RESULT]Recommended Action: {action}[/RESULT]\n"
yield f"[OBSERVATION]{justification}[/OBSERVATION]\n"
result = {
"action": action,
"justification": justification
}
return result
except Exception as e:
yield f"[ERROR]Failed to extract recommendation: {e}[/ERROR]\n"
return {"action": "UNKNOWN", "error": str(e)}
def compare_followup_images(
self,
previous_image_path: str,
current_image_path: str
) -> Generator[str, None, None]:
"""
Compare a follow-up image with the previous one.
Runs full analysis pipeline on current image, then compares findings.
"""
yield f"\n[STAGE:comparison]Follow-up Comparison Analysis[/STAGE]\n"
try:
current_image = Image.open(current_image_path).convert("RGB")
except Exception as e:
yield f"[ERROR]Failed to load images: {e}[/ERROR]\n"
return
# Store previous analysis state
prev_exam = self.last_medgemma_exam
# Generate comparison image and MONET deltas via MCP
yield f"\n[STAGE:current_analysis]Analyzing Current Image[/STAGE]\n"
if self.tools_loaded:
try:
compare_data = self.mcp_client.call_tool_sync(
"compare_images",
{
"image1_path": previous_image_path,
"image2_path": current_image_path,
},
)
yield f"[COMPARISON_IMAGE:{compare_data['comparison_path']}]\n"
# Side-by-side GradCAM comparison if both paths available
prev_gc = compare_data.get("prev_gradcam_path")
curr_gc = compare_data.get("curr_gradcam_path")
if prev_gc and curr_gc:
yield f"[GRADCAM_COMPARE:{prev_gc}:{curr_gc}]\n"
# Display MONET feature deltas
if compare_data["monet_deltas"]:
yield f"[TOOL_OUTPUT:Feature Comparison]\n"
for name, delta_info in compare_data["monet_deltas"].items():
prev_val = delta_info["previous"]
curr_val = delta_info["current"]
diff = delta_info["delta"]
short_name = name.replace("MONET_", "").replace("_", " ").title()
direction = "↑" if diff > 0 else "↓"
yield f" {short_name}: {prev_val:.0%} → {curr_val:.0%} ({direction}{abs(diff):.0%})\n"
yield f"[/TOOL_OUTPUT]\n"
except Exception as e:
yield f"[ERROR]MCP comparison failed: {e}[/ERROR]\n"
# MedGemma comparison analysis
comparison_prompt = f"""You are comparing TWO images of the same skin lesion taken at different times.
PREVIOUS ANALYSIS:
{prev_exam.get('synthesis', 'Not available')[:500] if prev_exam else 'Not available'}
Now examine the CURRENT image and compare to your memory of the previous findings.
Assess for changes in:
1. SIZE: Has the lesion grown, shrunk, or stayed the same?
2. COLOR: Any new colors appeared? Any colors faded?
3. SHAPE/SYMMETRY: Has the shape changed? More or less symmetric?
4. BORDERS: Sharper, more irregular, or unchanged?
5. STRUCTURES: New dermoscopic structures? Lost structures?
Provide your assessment:
- CHANGE_LEVEL: SIGNIFICANT_CHANGE / MINOR_CHANGE / STABLE / IMPROVED
- Specific changes observed
- Clinical recommendation based on changes"""
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": current_image},
{"type": "text", "text": comparison_prompt}
]
}
]
try:
yield f"[THINKING]Comparing current image to previous findings...[/THINKING]\n"
output = self.pipe(messages, max_new_tokens=200)
comparison_result = output[0]["generated_text"][-1]["content"]
yield f"[RESPONSE]\n"
words = comparison_result.split()
for i, word in enumerate(words):
time.sleep(0.02)
yield word + (" " if i < len(words) - 1 else "")
yield f"\n[/RESPONSE]\n"
# Extract change level
change_level = "UNKNOWN"
for level in ["SIGNIFICANT_CHANGE", "MINOR_CHANGE", "STABLE", "IMPROVED"]:
if level in comparison_result.upper():
change_level = level
break
if change_level == "SIGNIFICANT_CHANGE":
yield f"[RESULT]⚠️ SIGNIFICANT CHANGES DETECTED - Further evaluation recommended[/RESULT]\n"
elif change_level == "IMPROVED":
yield f"[RESULT]✓ LESION IMPROVED - Continue monitoring[/RESULT]\n"
elif change_level == "STABLE":
yield f"[RESULT]✓ LESION STABLE - Continue scheduled follow-up[/RESULT]\n"
else:
yield f"[RESULT]Minor changes noted - Clinical correlation recommended[/RESULT]\n"
except Exception as e:
yield f"[ERROR]Comparison analysis failed: {e}[/ERROR]\n"
yield f"\n[COMPLETE]Follow-up comparison complete[/COMPLETE]\n"
def chat(self, message: str, image_path: Optional[str] = None) -> str:
"""Simple chat interface"""
if not self.loaded:
self.load_model()
content = []
if image_path:
image = Image.open(image_path).convert("RGB")
content.append({"type": "image", "image": image})
content.append({"type": "text", "text": message})
messages = [{"role": "user", "content": content}]
output = self.pipe(messages, max_new_tokens=200)
return output[0]["generated_text"][-1]["content"]
def chat_followup(self, message: str) -> Generator[str, None, None]:
"""
Handle follow-up questions using the stored analysis context.
Uses the last analyzed image and diagnosis to provide contextual responses.
"""
if not self.loaded:
yield "[ERROR]Model not loaded[/ERROR]\n"
return
if not self.last_diagnosis or not self.last_image:
yield "[ERROR]No previous analysis context. Please analyze an image first.[/ERROR]\n"
return
# Build context from previous analysis
top_diagnosis = self.last_diagnosis['predictions'][0]
differentials = ", ".join([
f"{p['class']} ({p['probability']:.0%})"
for p in self.last_diagnosis['predictions'][:3]
])
monet_desc = ""
if self.last_monet_result:
monet_desc = self.last_monet_result.get('description', '')
context_prompt = f"""You are a dermatology assistant helping with skin lesion analysis.
PREVIOUS ANALYSIS CONTEXT:
- Primary diagnosis: {top_diagnosis['full_name']} ({top_diagnosis['probability']:.1%} confidence)
- Differential diagnoses: {differentials}
- Visual features: {monet_desc}
The user has a follow-up question about this lesion. Please provide a helpful, medically accurate response.
USER QUESTION: {message}
Provide a concise, informative response. If the question is outside your expertise or requires in-person examination, say so."""
messages = [
{
"role": "user",
"content": [
{"type": "image", "image": self.last_image},
{"type": "text", "text": context_prompt}
]
}
]
try:
yield f"[THINKING]Considering your question in context of the previous analysis...[/THINKING]\n"
time.sleep(0.2)
output = self.pipe(messages, max_new_tokens=200)
response = output[0]["generated_text"][-1]["content"]
yield f"[RESPONSE]\n"
# Stream word by word for typewriter effect
words = response.split()
for i, word in enumerate(words):
time.sleep(0.02)
yield word + (" " if i < len(words) - 1 else "")
yield f"\n[/RESPONSE]\n"
except Exception as e:
yield f"[ERROR]Failed to generate response: {e}[/ERROR]\n"
def main():
"""Interactive terminal interface"""
print("=" * 60)
print(" MedGemma Agent - Medical Image Analysis")
print("=" * 60)
agent = MedGemmaAgent(verbose=True)
agent.load_model()
print("\nCommands: analyze <path>, chat <message>, quit")
while True:
try:
user_input = input("\n> ").strip()
if not user_input:
continue
if user_input.lower() in ["quit", "exit", "q"]:
break
parts = user_input.split(maxsplit=1)
cmd = parts[0].lower()
if cmd == "analyze" and len(parts) > 1:
for chunk in agent.analyze_image_stream(parts[1].strip()):
print(chunk, end="", flush=True)
elif cmd == "chat" and len(parts) > 1:
print(agent.chat(parts[1]))
else:
print("Unknown command")
except KeyboardInterrupt:
break
except Exception as e:
print(f"Error: {e}")
if __name__ == "__main__":
main()
|