test-app / src /enhanced_ai_processor.py
SmartHeal's picture
Update src/enhanced_ai_processor.py
9b394b1 verified
import os
import logging
import cv2
import numpy as np
from PIL import Image
import torch
import json
from datetime import datetime
import tensorflow as tf
from transformers import pipeline
from ultralytics import YOLO
from tensorflow.keras.models import load_model
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from huggingface_hub import HfApi, HfFolder
import spaces
import time
from typing import Dict, Any, Optional, Tuple
from .config import Config
class EnhancedAIProcessor:
"""Enhanced AI processor with dashboard integration and analytics tracking"""
def __init__(self):
self.models_cache = {}
self.knowledge_base_cache = {}
self.config = Config()
self.px_per_cm = self.config.PIXELS_PER_CM
self.model_version = "v1.2.0" # Version for tracking
self._initialize_models()
@spaces.GPU(enable_queue=True, duration=90)
def _initialize_models(self):
"""Initialize all AI models including real-time models"""
try:
# Set HuggingFace token
if self.config.HF_TOKEN:
HfFolder.save_token(self.config.HF_TOKEN)
logging.info("HuggingFace token set successfully")
# Initialize MedGemma pipeline for medical text generation
try:
self.models_cache["medgemma_pipe"] = pipeline(
"image-text-to-text",
model="google/medgemma-4b-it",
torch_dtype=torch.bfloat16,
offload_folder="offload",
device_map="auto",
token=self.config.HF_TOKEN
)
logging.info("✅ MedGemma pipeline loaded successfully")
except Exception as e:
logging.warning(f"MedGemma pipeline not available: {e}")
# Initialize YOLO model for wound detection
try:
self.models_cache["det"] = YOLO(self.config.YOLO_MODEL_PATH)
logging.info("✅ YOLO detection model loaded successfully")
except Exception as e:
logging.warning(f"YOLO model not available: {e}")
# Initialize segmentation model
try:
self.models_cache["seg"] = load_model(self.config.SEG_MODEL_PATH, compile=False)
logging.info("✅ Segmentation model loaded successfully")
except Exception as e:
logging.warning(f"Segmentation model not available: {e}")
# Initialize wound classification model
try:
self.models_cache["cls"] = pipeline(
"image-classification",
model="Hemg/Wound-classification",
token=self.config.HF_TOKEN,
device="cpu"
)
logging.info("✅ Wound classification model loaded successfully")
except Exception as e:
logging.warning(f"Wound classification model not available: {e}")
# Initialize embedding model for knowledge base
try:
self.models_cache["embedding_model"] = HuggingFaceEmbeddings(
model_name="sentence-transformers/all-MiniLM-L6-v2",
model_kwargs={'device': 'cpu'}
)
logging.info("✅ Embedding model loaded successfully")
except Exception as e:
logging.warning(f"Embedding model not available: {e}")
logging.info("✅ All models loaded.")
self._load_knowledge_base()
except Exception as e:
logging.error(f"Error initializing AI models: {e}")
def _load_knowledge_base(self):
"""Load knowledge base from PDF guidelines"""
try:
documents = []
for pdf_path in self.config.GUIDELINE_PDFS:
if os.path.exists(pdf_path):
loader = PyPDFLoader(pdf_path)
docs = loader.load()
documents.extend(docs)
logging.info(f"Loaded PDF: {pdf_path}")
if documents and 'embedding_model' in self.models_cache:
# Split documents into chunks
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=1000,
chunk_overlap=100
)
chunks = text_splitter.split_documents(documents)
# Create vector store
vectorstore = FAISS.from_documents(chunks, self.models_cache['embedding_model'])
self.knowledge_base_cache['vectorstore'] = vectorstore
logging.info(f"✅ Knowledge base loaded with {len(chunks)} chunks")
else:
self.knowledge_base_cache['vectorstore'] = None
logging.warning("Knowledge base not available - no PDFs found or embedding model unavailable")
except Exception as e:
logging.warning(f"Knowledge base loading error: {e}")
self.knowledge_base_cache['vectorstore'] = None
def perform_comprehensive_analysis(self, image_pil: Image.Image, patient_info: Dict[str, Any]) -> Dict[str, Any]:
"""
Perform comprehensive analysis with enhanced tracking for dashboard integration
"""
start_time = time.time()
try:
# Perform visual analysis
visual_results = self.perform_visual_analysis(image_pil)
# Query guidelines for context
guideline_query = f"wound care {visual_results.get('wound_type', 'general')} treatment recommendations"
guideline_context = self.query_guidelines(guideline_query)
# Generate comprehensive report
report = self.generate_final_report(patient_info, visual_results, guideline_context, image_pil)
# Calculate processing time
processing_time = round(time.time() - start_time, 2)
# Calculate risk score based on multiple factors
risk_score = self._calculate_risk_score(visual_results, patient_info)
# Prepare comprehensive analysis data
analysis_data = {
'visual_results': visual_results,
'patient_info': patient_info,
'guideline_context': guideline_context,
'report': report,
'processing_time': processing_time,
'risk_score': risk_score,
'model_version': self.model_version,
'analysis_timestamp': datetime.now().isoformat(),
'analysis_metadata': {
'models_used': list(self.models_cache.keys()),
'image_dimensions': image_pil.size,
'guideline_sources': len(guideline_context.split('\n\n')) if guideline_context else 0
}
}
logging.info(f"✅ Comprehensive analysis completed in {processing_time}s with risk score {risk_score}")
return analysis_data
except Exception as e:
processing_time = round(time.time() - start_time, 2)
logging.error(f"❌ Analysis failed after {processing_time}s: {e}")
# Return error analysis data
return {
'error': str(e),
'processing_time': processing_time,
'risk_score': 0,
'model_version': self.model_version,
'analysis_timestamp': datetime.now().isoformat()
}
def perform_visual_analysis(self, image_pil: Image.Image) -> Dict[str, Any]:
"""Perform comprehensive visual analysis of wound image with enhanced tracking"""
try:
# Convert PIL to OpenCV format
image_cv = cv2.cvtColor(np.array(image_pil), cv2.COLOR_RGB2BGR)
# YOLO detection
if 'det' not in self.models_cache:
raise ValueError("YOLO detection model not available.")
results = self.models_cache['det'].predict(image_cv, verbose=False, device="cpu")
if not results or not results[0].boxes:
raise ValueError("No wound detected in the image.")
# Extract bounding box
box = results[0].boxes[0].xyxy[0].cpu().numpy().astype(int)
x1, y1, x2, y2 = box
region_cv = image_cv[y1:y2, x1:x2]
# Save detection image with timestamp
detection_image_cv = image_cv.copy()
cv2.rectangle(detection_image_cv, (x1, y1), (x2, y2), (0, 255, 0), 2)
os.makedirs(os.path.join(self.config.UPLOADS_DIR, "analysis"), exist_ok=True)
timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
detection_image_path = os.path.join(self.config.UPLOADS_DIR, "analysis", f"detection_{timestamp}.png")
cv2.imwrite(detection_image_path, detection_image_cv)
detection_image_pil = Image.fromarray(cv2.cvtColor(detection_image_cv, cv2.COLOR_BGR2RGB))
# Initialize outputs
length = breadth = area = 0
segmentation_image_pil = None
segmentation_image_path = None
segmentation_confidence = 0.0
# Segmentation (optional)
if 'seg' in self.models_cache:
input_size = self.models_cache['seg'].input_shape[1:3] # (height, width)
resized_region = cv2.resize(region_cv, (input_size[1], input_size[0]))
seg_input = np.expand_dims(resized_region / 255.0, 0)
mask_pred = self.models_cache['seg'].predict(seg_input, verbose=0)[0]
mask_np = (mask_pred[:, :, 0] > 0.5).astype(np.uint8)
# Calculate segmentation confidence
segmentation_confidence = float(np.mean(mask_pred[:, :, 0]))
# Resize mask back to original region size
mask_resized = cv2.resize(mask_np, (region_cv.shape[1], region_cv.shape[0]), interpolation=cv2.INTER_NEAREST)
# Overlay mask on region for visualization
overlay = region_cv.copy()
overlay[mask_resized == 1] = [0, 0, 255] # Red overlay
# Blend overlay for final output
segmented_visual = cv2.addWeighted(region_cv, 0.7, overlay, 0.3, 0)
# Save segmentation image
segmentation_image_path = os.path.join(self.config.UPLOADS_DIR, "analysis", f"segmentation_{timestamp}.png")
cv2.imwrite(segmentation_image_path, segmented_visual)
segmentation_image_pil = Image.fromarray(cv2.cvtColor(segmented_visual, cv2.COLOR_BGR2RGB))
# Wound measurements from resized mask
contours, _ = cv2.findContours(mask_resized, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if contours:
cnt = max(contours, key=cv2.contourArea)
x, y, w, h = cv2.boundingRect(cnt)
length = round(h / self.px_per_cm, 2)
breadth = round(w / self.px_per_cm, 2)
area = round(cv2.contourArea(cnt) / (self.px_per_cm ** 2), 2)
# Classification with confidence tracking
wound_type = "Unknown"
classification_confidence = 0.0
classification_scores = []
if 'cls' in self.models_cache:
try:
region_pil = Image.fromarray(cv2.cvtColor(region_cv, cv2.COLOR_BGR2RGB))
cls_result = self.models_cache['cls'](region_pil)
if cls_result:
best_result = max(cls_result, key=lambda x: x['score'])
wound_type = best_result['label']
classification_confidence = float(best_result['score'])
classification_scores = [{'label': r['label'], 'score': float(r['score'])} for r in cls_result]
except Exception as e:
logging.warning(f"Wound classification error: {e}")
return {
'wound_type': wound_type,
'length_cm': length,
'breadth_cm': breadth,
'surface_area_cm2': area,
'detection_confidence': float(results[0].boxes[0].conf.cpu().item()),
'segmentation_confidence': segmentation_confidence,
'classification_confidence': classification_confidence,
'classification_scores': classification_scores,
'bounding_box': box.tolist(),
'detection_image_path': detection_image_path,
'detection_image_pil': detection_image_pil,
'segmentation_image_path': segmentation_image_path,
'segmentation_image_pil': segmentation_image_pil,
'analysis_quality': {
'detection_quality': 'high' if float(results[0].boxes[0].conf.cpu().item()) > 0.8 else 'medium',
'segmentation_quality': 'high' if segmentation_confidence > 0.7 else 'medium',
'classification_quality': 'high' if classification_confidence > 0.8 else 'medium'
}
}
except Exception as e:
logging.error(f"Visual analysis error: {e}")
raise ValueError(f"Visual analysis failed: {str(e)}")
def _calculate_risk_score(self, visual_results: Dict[str, Any], patient_info: Dict[str, Any]) -> int:
"""
Calculate comprehensive risk score (0-100) based on visual analysis and patient data
"""
try:
risk_score = 0
# Wound size risk (0-25 points)
area = visual_results.get('surface_area_cm2', 0)
if area > 10:
risk_score += 25
elif area > 5:
risk_score += 15
elif area > 2:
risk_score += 10
else:
risk_score += 5
# Wound type risk (0-20 points)
wound_type = visual_results.get('wound_type', '').lower()
high_risk_types = ['ulcer', 'necrotic', 'infected', 'diabetic']
medium_risk_types = ['pressure', 'venous', 'arterial']
if any(risk_type in wound_type for risk_type in high_risk_types):
risk_score += 20
elif any(risk_type in wound_type for risk_type in medium_risk_types):
risk_score += 15
else:
risk_score += 10
# Patient factors (0-30 points)
age = patient_info.get('patient_age', 0)
if age > 70:
risk_score += 15
elif age > 50:
risk_score += 10
else:
risk_score += 5
# Diabetic status
diabetic_status = patient_info.get('diabetic_status', '').lower()
if 'yes' in diabetic_status or 'diabetic' in diabetic_status:
risk_score += 15
# Pain level (0-10 points)
pain_level = patient_info.get('pain_level', 0)
if pain_level > 7:
risk_score += 10
elif pain_level > 4:
risk_score += 7
else:
risk_score += 3
# Infection signs (0-15 points)
infection_signs = patient_info.get('infection_signs', '').lower()
if 'yes' in infection_signs or 'present' in infection_signs:
risk_score += 15
elif 'possible' in infection_signs or 'mild' in infection_signs:
risk_score += 10
else:
risk_score += 5
# Ensure score is within 0-100 range
risk_score = min(max(risk_score, 0), 100)
logging.info(f"Calculated risk score: {risk_score}")
return risk_score
except Exception as e:
logging.error(f"Error calculating risk score: {e}")
return 50 # Default medium risk
def query_guidelines(self, query: str) -> str:
"""Query the knowledge base for relevant guidelines with enhanced tracking"""
try:
vector_store = self.knowledge_base_cache.get("vectorstore")
if not vector_store:
return "Knowledge base unavailable - clinical guidelines not loaded"
# Retrieve relevant documents
retriever = vector_store.as_retriever(search_kwargs={"k": 10})
docs = retriever.invoke(query)
if not docs:
return "No relevant guidelines found for the query"
# Format the results with enhanced metadata
formatted_results = []
for i, doc in enumerate(docs):
source = doc.metadata.get('source', 'Unknown')
page = doc.metadata.get('page', 'N/A')
content = doc.page_content.strip()
# Add relevance indicator
relevance = f"Result {i+1}/10"
formatted_results.append(f"[{relevance}] Source: {source}, Page: {page}\nContent: {content}")
guideline_text = "\n\n".join(formatted_results)
logging.info(f"Retrieved {len(docs)} guideline documents for query: {query[:50]}...")
return guideline_text
except Exception as e:
logging.error(f"Guidelines query error: {e}")
return f"Error querying guidelines: {str(e)}"
@spaces.GPU(enable_queue=True, duration=90)
def generate_final_report(self, patient_info: Dict[str, Any], visual_results: Dict[str, Any],
guideline_context: str, image_pil: Image.Image, max_new_tokens: int = None) -> str:
"""Generate comprehensive medical report using MedGemma with enhanced tracking"""
try:
if 'medgemma_pipe' not in self.models_cache:
return self._generate_fallback_report(patient_info, visual_results, guideline_context)
max_tokens = max_new_tokens or self.config.MAX_NEW_TOKENS
# Get detection and segmentation images if available
detection_image = visual_results.get('detection_image_pil', None)
segmentation_image = visual_results.get('segmentation_image_pil', None)
# Create enhanced prompt with quality indicators
analysis_quality = visual_results.get('analysis_quality', {})
prompt = f"""
# SmartHeal AI Wound Care Report
## Patient Information
{self._format_patient_info(patient_info)}
## Visual Analysis Summary
- Wound Type: {visual_results.get('wound_type', 'Unknown')} (Confidence: {visual_results.get('classification_confidence', 0):.2f})
- Dimensions: {visual_results.get('length_cm', 0)} × {visual_results.get('breadth_cm', 0)} cm
- Surface Area: {visual_results.get('surface_area_cm2', 0)} cm²
- Detection Quality: {analysis_quality.get('detection_quality', 'medium')}
- Segmentation Quality: {analysis_quality.get('segmentation_quality', 'medium')}
## Clinical Reference Guidelines
{guideline_context[:2000]}...
## Analysis Request
You are SmartHeal-AI Agent, a specialized wound care AI with expertise in clinical assessment and evidence-based treatment planning.
Based on the comprehensive data provided (patient information, precise wound measurements, clinical guidelines, and visual analysis), generate a structured clinical report with the following sections:
### 1. Clinical Assessment
- Detailed wound characterization based on visual analysis
- Tissue type assessment (granulation, slough, necrotic, epithelializing)
- Peri-wound skin condition evaluation
- Infection risk assessment
### 2. Treatment Recommendations
- Specific wound care dressing recommendations based on wound characteristics
- Topical treatments if indicated
- Debridement recommendations if needed
- Pressure offloading strategies if applicable
### 3. Risk Stratification
- Patient-specific risk factors analysis
- Healing prognosis assessment
- Complications to monitor
### 4. Follow-up Plan
- Recommended assessment frequency
- Key monitoring parameters
- Escalation criteria for specialist referral
Generate a concise, evidence-based report suitable for clinical documentation.
"""
# Prepare messages for MedGemma with all available images
content_list = [{"type": "text", "text": prompt}]
# Add images in order of importance
if image_pil:
content_list.insert(0, {"type": "image", "image": image_pil})
if detection_image:
content_list.insert(1, {"type": "image", "image": detection_image})
if segmentation_image:
content_list.insert(2, {"type": "image", "image": segmentation_image})
messages = [
{
"role": "system",
"content": [{"type": "text", "text": "You are a specialized medical AI assistant for wound care with expertise in clinical assessment, treatment planning, and evidence-based recommendations. Provide structured, actionable clinical reports."}],
},
{
"role": "user",
"content": content_list
}
]
# Generate report using MedGemma
output = self.models_cache['medgemma_pipe'](
text=messages,
max_new_tokens=max_tokens,
do_sample=False,
)
generated_content = output[0]['generated_text']
# Extract the assistant's response
if isinstance(generated_content, list):
for message in generated_content:
if message.get('role') == 'assistant':
report_content = message.get('content', '')
if isinstance(report_content, list):
report_text = ''.join([item.get('text', '') for item in report_content if item.get('type') == 'text'])
else:
report_text = str(report_content)
break
else:
report_text = str(generated_content)
else:
report_text = str(generated_content)
# Add metadata to report
report_with_metadata = f"""
{report_text}
---
**Report Metadata:**
- Generated by: SmartHeal AI v{self.model_version}
- Analysis Quality: Detection ({analysis_quality.get('detection_quality', 'medium')}), Segmentation ({analysis_quality.get('segmentation_quality', 'medium')})
- Generated at: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
"""
logging.info("✅ MedGemma report generated successfully")
return report_with_metadata
except Exception as e:
logging.error(f"MedGemma report generation error: {e}")
return self._generate_fallback_report(patient_info, visual_results, guideline_context)
def _format_patient_info(self, patient_info: Dict[str, Any]) -> str:
"""Format patient information for report"""
formatted = f"""
- Name: {patient_info.get('patient_name', 'N/A')}
- Age: {patient_info.get('patient_age', 'N/A')} years
- Gender: {patient_info.get('patient_gender', 'N/A')}
- Wound Location: {patient_info.get('wound_location', 'N/A')}
- Wound Duration: {patient_info.get('wound_duration', 'N/A')}
- Pain Level: {patient_info.get('pain_level', 'N/A')}/10
- Diabetic Status: {patient_info.get('diabetic_status', 'N/A')}
- Infection Signs: {patient_info.get('infection_signs', 'N/A')}
- Previous Treatment: {patient_info.get('previous_treatment', 'N/A')}
- Medical History: {patient_info.get('medical_history', 'N/A')}
- Current Medications: {patient_info.get('medications', 'N/A')}
- Known Allergies: {patient_info.get('allergies', 'N/A')}
"""
return formatted.strip()
def _generate_fallback_report(self, patient_info: Dict[str, Any], visual_results: Dict[str, Any],
guideline_context: str) -> str:
"""Generate fallback report when MedGemma is not available"""
wound_type = visual_results.get('wound_type', 'Unknown')
length = visual_results.get('length_cm', 0)
breadth = visual_results.get('breadth_cm', 0)
area = visual_results.get('surface_area_cm2', 0)
# Basic risk assessment
risk_factors = []
if patient_info.get('patient_age', 0) > 65:
risk_factors.append("Advanced age")
if 'yes' in str(patient_info.get('diabetic_status', '')).lower():
risk_factors.append("Diabetes mellitus")
if patient_info.get('pain_level', 0) > 6:
risk_factors.append("High pain level")
if area > 5:
risk_factors.append("Large wound size")
report = f"""
# SmartHeal AI Wound Assessment Report
## Clinical Summary
**Patient:** {patient_info.get('patient_name', 'N/A')}, {patient_info.get('patient_age', 'N/A')} years old {patient_info.get('patient_gender', '')}
**Wound Characteristics:**
- Type: {wound_type}
- Location: {patient_info.get('wound_location', 'N/A')}
- Dimensions: {length} × {breadth} cm (Area: {area} cm²)
- Duration: {patient_info.get('wound_duration', 'N/A')}
- Pain Level: {patient_info.get('pain_level', 'N/A')}/10
## Risk Assessment
**Identified Risk Factors:**
{chr(10).join(f'- {factor}' for factor in risk_factors) if risk_factors else '- No significant risk factors identified'}
## Treatment Recommendations
**Wound Care:**
- Regular wound assessment and documentation
- Appropriate dressing selection based on wound characteristics
- Maintain moist wound environment
- Monitor for signs of infection
**Patient Management:**
- Pain management as indicated
- Nutritional assessment and optimization
- Patient education on wound care
## Follow-up Plan
- Reassess wound in 1-2 weeks
- Monitor for signs of healing or deterioration
- Consider specialist referral if no improvement in 4 weeks
---
**Report Generated by:** SmartHeal AI Fallback System v{self.model_version}
**Generated at:** {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}
**Note:** This is a basic assessment. For comprehensive analysis, ensure all AI models are properly loaded.
"""
logging.info("✅ Fallback report generated")
return report