app.py

import gradio as gr
import torch
from PIL import Image
from transformers import (
    CLIPProcessor, CLIPModel,
    BlipProcessor, BlipForConditionalGeneration,
    GPT2Tokenizer, GPT2LMHeadModel
)
from unsloth import FastLanguageModel  # For fast quantized TinyLlama

# Device and dtype setup
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = None if torch.cuda.is_available() else torch.float32

# Load models (Gradio/HF caches 'em)
@gr.cache
def load_models():
    clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32", torch_dtype=torch.float16 if device == "cuda" else torch.float32).to(device)
    clip_proc = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32", use_fast=False)
    blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base", torch_dtype=torch.float16 if device == "cuda" else torch.float32).to(device)
    blip_proc = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base", use_fast=True)
    
    # Unsloth for fast TinyLlama (fallback to regular if issues)
    try:
        model, llama_tok = FastLanguageModel.from_pretrained(
            model_name="unsloth/tinyllama-bnb-4bit",
            max_seq_length=2048,
            dtype=dtype,
            load_in_4bit=True,
        )
        FastLanguageModel.for_inference(model)
    except:
        # Fallback to regular
        from transformers import AutoTokenizer, AutoModelForCausalLM
        llama_tok = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
        if llama_tok.pad_token is None:
            llama_tok.pad_token = llama_tok.eos_token
        model = AutoModelForCausalLM.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0", torch_dtype=torch.float16 if device == "cuda" else torch.float32).to(device)
    
    gpt2_tok = GPT2Tokenizer.from_pretrained("distilgpt2")
    if gpt2_tok.pad_token is None:
        gpt2_tok.pad_token = gpt2_tok.eos_token
    gpt2_model = GPT2LMHeadModel.from_pretrained("distilgpt2", torch_dtype=torch.float16 if device == "cuda" else torch.float32).to(device)
    
    return clip_model, clip_proc, blip_model, blip_proc, llama_tok, model, gpt2_tok, gpt2_model

clip_model, clip_proc, blip_model, blip_proc, llama_tok, llama_model, gpt2_tok, gpt2_model = load_models()

def generate_report(image):
    if image is None:
        return "Upload an X-ray to get started!", "", ""
    
    image = image.convert("RGB").resize((224, 224))
    
    # Step 1: Caption
    with torch.no_grad():
        blip_inputs = blip_proc(images=image, return_tensors="pt").to(device)
        caption_ids = blip_model.generate(**blip_inputs)
        caption = blip_proc.decode(caption_ids[0], skip_special_tokens=True)
    
    # Step 2: Findings
    reference_findings = [
        "There is evidence of right lower lobe consolidation.",
        "No acute cardiopulmonary abnormality.",
        "Mild cardiomegaly with clear lungs.",
        "Findings consistent with pneumonia.",
        "Chronic interstitial changes noted."
    ]
    with torch.no_grad():
        img_inputs = clip_proc(images=image, return_tensors="pt").to(device)
        img_features = clip_model.get_image_features(**img_inputs)
        img_features = torch.nn.functional.normalize(img_features, dim=-1)
        
        text_inputs = clip_proc.tokenizer(reference_findings, return_tensors="pt", padding=True, truncation=True).to(device)
        txt_features = clip_model.get_text_features(**text_inputs)
        txt_features = torch.nn.functional.normalize(txt_features, dim=-1)
        
        similarities = torch.matmul(txt_features, img_features.T).squeeze()
        top_indices = torch.topk(similarities, k=3).indices
        top_findings = [reference_findings[i] for i in top_indices]
    
    findings_text = f"• {top_findings[0]}\n• {top_findings[1]}\n• {top_findings[2]}"
    
    # Step 3: Draft
    llama_prompt = f"🧠 Caption: {caption}\n📚 Retrieved Reports:\n" + "\n".join(f"- {f}" for f in top_findings) + "\n\nGenerate a clinical-style radiology report:"
    with torch.no_grad():
        inputs = llama_tok(llama_prompt, return_tensors="pt").to(device)
        outputs = llama_model.generate(**inputs, max_new_tokens=75, use_cache=True, do_sample=True, temperature=0.7)
        draft_report = llama_tok.decode(outputs[0], skip_special_tokens=True)
    
    # Steps 4-5: Refine
    gpt2_input_1 = f"Caption: {caption}\nDraft: {draft_report}\nRefine this into a structured radiology report:"
    with torch.no_grad():
        gpt2_inputs_1 = gpt2_tok(gpt2_input_1, return_tensors="pt").to(device)
        gpt2_output_1 = gpt2_model.generate(**gpt2_inputs_1, max_new_tokens=100, do_sample=True, temperature=0.7, pad_token_id=gpt2_tok.eos_token_id)
        refined_1 = gpt2_tok.decode(gpt2_output_1[0], skip_special_tokens=True)
        
        gpt2_input_2 = f"{refined_1}\nRefine the new report:"
        gpt2_inputs_2 = gpt2_tok(gpt2_input_2, return_tensors="pt").to(device)
        gpt2_output_2 = gpt2_model.generate(**gpt2_inputs_2, max_new_tokens=75, do_sample=True, temperature=0.7, pad_token_id=gpt2_tok.eos_token_id)
        refined_2 = gpt2_tok.decode(gpt2_output_2[0], skip_special_tokens=True)
    
    return f"**Caption:** {caption}", findings_text, refined_2

# Gradio Interface
with gr.Blocks(title="🩺 AI Radiology Assistant") as demo:
    gr.Markdown("# 🩺 AI-Powered Chest X-ray Interpretation Tool")
    
    with gr.Row():
        image_input = gr.Image(type="pil", label="Upload Chest X-ray (PNG/JPG)")
    
    caption_out = gr.Textbox(label="🧠 AI Captioning", interactive=False)
    findings_out = gr.Textbox(label="🔍 Top Similar Clinical Findings", interactive=False)
    report_out = gr.Textbox(label="📄 Final Structured Radiology Report", lines=20, interactive=False)
    
    submit_btn = gr.Button("Generate Report", variant="primary")
    
    submit_btn.click(
        generate_report,
        inputs=image_input,
        outputs=[caption_out, findings_out, report_out]
    )

if __name__ == "__main__":
    demo.launch()