app_2.py

from fastai.vision.all import *
import gradio as gr
from pathlib import Path
import pandas as pd

def get_x(row):
    # All files are assumed to be '.jpg', so we directly return the path with '.jpg' extension
    return path_image_combined / f"{row['file_name']}.jpg"

def get_y(row):
    return row['Buried ODD']

# Load the models into a dictionary
models = {
    'Ultrasound': load_learner('ODDUltrasound.pkl'),
    'OCT': load_learner('ODDOCT.pkl'),
    'Fundus': load_learner('ODDfundus.pkl'),
    'Fluorescence': load_learner('ODDfluorescence.pkl')
}

modality_keys = ['Ultrasound', 'OCT', 'Fundus', 'Fluorescence']

def classify_images(img_ultrasound, img_oct, img_fundus, img_fluorescence):
    imgs = [img_ultrasound, img_oct, img_fundus, img_fluorescence]
    predictions = []
    detailed_predictions = []  # To store detailed predictions for each modality
    
    # Convert images to PILImage and predict with each model
    for img, key in zip(imgs, modality_keys):
        pil_img = PILImage.create(img)
        pred, _, probs = models[key].predict(pil_img)
        predictions.append(pred)
        
        # Assuming binary classification, extract the probability for the predicted class
        prob_pred = probs.max()  # Get the highest probability score
        detailed_predictions.append(f"{key}: {pred} ({prob_pred:.2f}%)")  # Format prediction with percentage
    
    # Majority vote for final decision
    final_decision = max(set(predictions), key=predictions.count)
    
    detailed_predictions.append(f"Final Decision: {final_decision}")  # Add the final decision
    
    return "\n".join(detailed_predictions)  # Return detailed predictions as a single string

# Define the Gradio interface inputs and outputs
inputs = [gr.Image(label=f"{modality} Image") for modality in modality_keys]
output = gr.Text(label="Predictions")

intf = gr.Interface(fn=classify_images, inputs=inputs, outputs=output,
                    title="ODD Detection from Multiple Imaging Modalities",
                    description="Upload images for each modality and receive individual predictions with percentages and a binary prediction for Optic Disk Drusen presence.")
intf.launch(share=True)