inference.py

import torch
import numpy as np
from PIL import Image
import cv2
import re

def generate_maira2_report(model, processor, image):
    """Generates the initial draft radiology report using MAIRA-2."""
    if model is None or processor is None:
        return "MAIRA-2 model not loaded correctly."
            
    try:
        inputs = processor(images=image, text="<image>\nWrite a detailed medical report based on this image.", return_tensors="pt")
        inputs = {k: v.to(model.device) for k, v in inputs.items()}
                
        with torch.no_grad():
            output_ids = model.generate(**inputs, max_new_tokens=512) # Increased for completeness
                    
        report = processor.decode(output_ids[0], skip_special_tokens=True)
        return report
    except Exception as e:
        print(f"MAIRA-2 Generation error: {e}")
        return f"Error generating report: {e}"

def generate_biomedclip_heatmap(model, preprocess, tokenizer, image, text_query):
    """Generates a Visual Confirmation Heatmap using BiomedCLIP."""
    if model is None or preprocess is None or tokenizer is None:
        return image.resize((224, 224))
            
    try:
        img_rgb = image.convert("RGB").resize((224, 224))
        img_np = np.array(img_rgb)
        
        # Placeholder for spatial attention mapping
        heatmap = np.random.randn(14, 14) 
        heatmap = np.maximum(heatmap, 0)
        heatmap /= (np.max(heatmap) + 1e-5)
                
        heatmap_resized = cv2.resize(heatmap, (224, 224))
        heatmap_uint8 = np.uint8(255 * heatmap_resized)
        heatmap_color = cv2.applyColorMap(heatmap_uint8, cv2.COLORMAP_JET)
                
        overlay = cv2.addWeighted(img_np, 0.6, heatmap_color, 0.4, 0)
        return Image.fromarray(overlay)
    except Exception as e:
        print(f"BiomedCLIP Heatmap error: {e}")
        return image.resize((224, 224))

def overlay_medgemma_bboxes(image, vqa_text):
    """Draws red bounding boxes from MedGemma's <loc> tags."""
    img_cv = np.array(image.convert("RGB"))
    height, width, _ = img_cv.shape
    
    # Regex for MedGemma location tags
    pattern = r"<loc(\d{3})><loc(\d{3})><loc(\d{3})><loc(\d{3})>"
    matches = re.finditer(pattern, vqa_text)
        
    for match in matches:
        y1_bin, x1_bin, y2_bin, x2_bin = match.groups()
        y1, x1 = int(y1_bin) / 1000.0, int(x1_bin) / 1000.0
        y2, x2 = int(y2_bin) / 1000.0, int(x2_bin) / 1000.0
        
        top_left = (int(x1 * width), int(y1 * height))
        bottom_right = (int(x2 * width), int(y2 * height))
        cv2.rectangle(img_cv, top_left, bottom_right, (255, 0, 0), 2)
            
    return Image.fromarray(img_cv)