app.py

import gradio as gr
from huggingface_hub import hf_hub_download
import torch
import torch.nn.functional as F
from clinicadl.utils.network.cnn.models import Conv5_FC3
import nibabel as nib 
import numpy as np

# Download model from Hub
model_path = hf_hub_download(repo_id="ARAMIS-LAB/CNN-AD-CN", filename="model.pth.tar")

# Load ClinicaDL model
checkpoint_state = torch.load(model_path, map_location="cpu")
model = Conv5_FC3(input_size= [
        1,
        169,
        208,
        179
    ], gpu = False)
model.load_state_dict(checkpoint_state["model"])
model.eval()

# Class labels
CLASSES = ["CN", "AD"]

def preprocess_nii(nii_file):
    # Load NIfTI file
    img = nib.load(nii_file)
    data = img.get_fdata()  # numpy array (float64)

    # Normalize intensities
    data = (data - np.mean(data)) / np.std(data)

    # Convert to tensor
    tensor = torch.tensor(data, dtype=torch.float32).unsqueeze(0).unsqueeze(0)  
    # Shape: [1, 1, D, H, W]

    # Resize or pad to expected input shape
    target_shape = (1, 1, 169, 208, 179)
    tensor = F.interpolate(tensor, size=target_shape[2:], mode="trilinear", align_corners=False)

    return tensor
    
def predict(input_image):
    x = preprocess_nii(input_image)
    with torch.no_grad():
        logits = model(x)
        probs = F.softmax(logits, dim=1)  # convert to probabilities
        
    probs = probs.squeeze(0).cpu().numpy().tolist()  # ensures it's a Python list of floats
    results = {cls: prob for cls, prob in zip(CLASSES, probs)}
    
    return results

# Gradio app: file upload instead of image
demo = gr.Interface(
    fn=predict,
    inputs=gr.File(type="filepath", label=".nii.gz MRI upload"),
    outputs="json",
    title="ClinicaDL MRI Classifier",
    description="Upload a .nii.gz file to get the model's prediction."
)

demo.launch()