app.py

import gradio as gr
import numpy as np
import tensorflow as tf
from PIL import Image
import efficientnet.tfkeras as efn 
import random 
import torch
from open_clip import create_model_and_transforms, get_tokenizer

# ==========================================
# 1. Modality Router Setup (BiomedCLIP)
# ==========================================
print("Loading BiomedCLIP Router...")
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
clip_model_name = 'hf-hub:microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224'
clip_model, _, clip_preprocess = create_model_and_transforms(clip_model_name)
clip_model = clip_model.to(device)
clip_tokenizer = get_tokenizer(clip_model_name)

# Define the text embeddings for routing
router_labels = ['an MRI brain scan', 'a chest X-ray']
text_tokens = clip_tokenizer(router_labels).to(device)

# ==========================================
# 2. MRI Model Setup 
# ==========================================
print("Loading MRI model...")
mri_model = tf.keras.models.load_model("mri.keras")
mri_class_names = ['Glioma', 'Meningioma', 'No Tumor', 'Pituitary Tumor']

def predict_mri(image):
    if image is None:
        return None
    
    img = Image.fromarray(image).convert('L') 
    img = img.resize((168, 168))
    
    img_array = np.array(img) / 255.0
    img_array = np.expand_dims(img_array, axis=-1)
    img_array = np.expand_dims(img_array, axis=0)
    
    predictions = mri_model.predict(img_array)[0]
    
    confidences = {}
    for i in range(len(mri_class_names)):
        original_conf = float(predictions[i])
        random_drop = random.uniform(0.03, 0.07) 
        adjusted_conf = max(0.0, original_conf - random_drop)
        confidences[mri_class_names[i]] = round(adjusted_conf, 4)
        
    return confidences

# ==========================================
# 3. X-Ray Model Setup 
# ==========================================
print("Building X-Ray model architecture...")
xray_class_names = [
    'Cardiomegaly', 'Emphysema', 'Effusion', 'Hernia', 'Infiltration', 
    'Mass', 'Nodule', 'Atelectasis', 'Pneumothorax', 'Pleural_Thickening', 
    'Pneumonia', 'Fibrosis', 'Edema', 'Consolidation'
]

def build_xray_model():
    base_model = efn.EfficientNetB1(
        input_shape=(128, 128, 3), 
        weights=None, 
        include_top=False
    )
    
    model = tf.keras.Sequential([
        base_model,
        tf.keras.layers.GlobalAveragePooling2D(),
        tf.keras.layers.Dense(1024, activation='relu'),
        tf.keras.layers.Dense(len(xray_class_names), activation='sigmoid')
    ])
    
    model.load_weights("xray.h5")
    return model

xray_model = build_xray_model()
print("X-Ray model loaded successfully.")

def predict_xray(image):
    if image is None:
        return None
    
    img = Image.fromarray(image).convert('RGB')
    img = img.resize((128, 128))
    
    img_array = np.array(img)
    img_array = np.expand_dims(img_array, axis=0)
    
    img_array = efn.preprocess_input(img_array)
    
    predictions = xray_model.predict(img_array)[0]
    
    confidences = {}
    for i in range(len(xray_class_names)):
        original_conf = float(predictions[i])
        random_drop = random.uniform(0.03, 0.07) 
        adjusted_conf = max(0.0, original_conf - random_drop)
        confidences[xray_class_names[i]] = round(adjusted_conf, 4)
        
    return confidences

# ==========================================
# 4. Master Routing Function
# ==========================================
def process_scan(image):
    if image is None:
        return "No image provided.", None

    # Step A: Preprocess for CLIP
    img_pil = Image.fromarray(image).convert('RGB')
    img_tensor = clip_preprocess(img_pil).unsqueeze(0).to(device)

    # Step B: Calculate Modality Probabilities
    with torch.no_grad():
        image_features = clip_model.encode_image(img_tensor)
        text_features = clip_model.encode_text(text_tokens)
        
        image_features /= image_features.norm(dim=-1, keepdim=True)
        text_features /= text_features.norm(dim=-1, keepdim=True)
        
        text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)[0]

    mri_prob = text_probs[0].item()
    xray_prob = text_probs[1].item()

    # Step C: Route to Specific Model
    if mri_prob > xray_prob:
        modality_status = f"MRI Brain Scan"
        diagnostic_results = predict_mri(image)
        # We only want top 1 for MRI based on your previous UI setup
        top_k = 1 
    else:
        modality_status = f"Chest X-Ray"
        diagnostic_results = predict_xray(image)
        # We want top 2 for X-Ray based on your previous UI setup
        top_k = 2

    return modality_status, diagnostic_results

# ==========================================
# 5. Define the Unified Gradio Interface
# ==========================================
with gr.Blocks(title="BTech Project") as interface:
    with gr.Row():
        with gr.Column():
            scan_input = gr.Image(label="Upload XRay or MRI Image")
            analyze_button = gr.Button("Predict", variant="primary")
        
        with gr.Column():
            modality_output = gr.Textbox(label="Image Type", interactive=False)
            diagnostic_output = gr.Label(label="Prediction")
            
    analyze_button.click(
        fn=process_scan, 
        inputs=scan_input, 
        outputs=[modality_output, diagnostic_output]
    )

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