app.py

import gradio as gr
import subprocess
import os
import shutil
from huggingface_hub import hf_hub_download
import torch
import nibabel as nib
import matplotlib.pyplot as plt
import spaces  # Import spaces for GPU decoration
import numpy as np
from scipy.ndimage import center_of_mass, zoom

# Define paths
MODEL_DIR = "./model"  # Local directory to store the downloaded model
DATASET_DIR = os.path.join(MODEL_DIR, "Dataset004_WML")  # Directory for Dataset004_WML
INPUT_DIR = "/tmp/input"
OUTPUT_DIR = "/tmp/output"

# Hugging Face Model Repository
REPO_ID = "FrancescoLR/FLAMeS-model"  # Replace with your actual model repository ID

# Function to download the Dataset004_WML folder
def download_model():
    if not os.path.exists(DATASET_DIR):
        os.makedirs(DATASET_DIR, exist_ok=True)
        print("Downloading Dataset004_WML.zip...")
        zip_path = hf_hub_download(repo_id=REPO_ID, filename="Dataset004_WML.zip", cache_dir=MODEL_DIR)
        subprocess.run(["unzip", "-o", zip_path, "-d", MODEL_DIR])
        print("Dataset004_WML downloaded and extracted.")
        
def resample_to_isotropic(data, affine, target_spacing=1.0):
    """
    Resamples a 3D NIfTI image to isotropic voxel size.
    
    Parameters:
        data (numpy.ndarray): The input 3D image data.
        affine (numpy.ndarray): The affine transformation matrix.
        target_spacing (float): Desired isotropic voxel spacing (in mm).
    
    Returns:
        resampled_data (numpy.ndarray): Resampled image data.
        resampled_affine (numpy.ndarray): Updated affine matrix.
    """
    # Extract current voxel dimensions from the affine matrix
    current_spacing = np.sqrt((affine[:3, :3] ** 2).sum(axis=0))

    # Compute the scaling factors for resampling
    scaling_factors = current_spacing / target_spacing

    # Resample the data using zoom
    resampled_data = zoom(data, zoom=scaling_factors, order=1)  # Linear interpolation

    # Update the affine matrix to reflect the new voxel dimensions
    resampled_affine = affine.copy()
    resampled_affine[:3, :3] /= scaling_factors[:, np.newaxis]

    return resampled_data, resampled_affine        
    
def extract_middle_slices(nifti_path, output_image_path, slice_size=180):
    """
    Extracts slices centered around the center of mass of non-zero voxels in a 3D NIfTI image.
    The slices are taken along axial, coronal, and sagittal planes and saved as a single PNG.
    """
  # Load NIfTI image
    img = nib.load(nifti_path)
    data = img.get_fdata()
    affine = img.affine

    # Resample the image to 1 mm isotropic
    resampled_data, _ = resample_to_isotropic(data, affine, target_spacing=1.0)

    # Compute the center of mass of non-zero voxels
    com = center_of_mass(resampled_data > 0)
    center = np.round(com).astype(int)

    # Define half the slice size
    half_size = slice_size // 2

def extract_middle_slices(nifti_path, output_image_path, slice_size=180, center=None):
    """
    Extracts slices from a 3D NIfTI image. If a center is provided, it uses it;
    otherwise, computes the center of mass of non-zero voxels. Slices are taken 
    along axial, coronal, and sagittal planes and saved as a single PNG.
    """
    # Load NIfTI image
    img = nib.load(nifti_path)
    data = img.get_fdata()
    affine = img.affine

    # Resample the image to 1 mm isotropic
    resampled_data, _ = resample_to_isotropic(data, affine, target_spacing=1.0)

    # Compute or reuse the center of mass
    if center is None:
        com = center_of_mass(resampled_data > 0)
        center = np.round(com).astype(int)

    # Define half the slice size
    half_size = slice_size // 2

    # Safely extract and pad 2D slices
    def extract_2d_slice(data, center, axis):
        slices = [slice(None)] * 3
        slices[axis] = center[axis]  # Fix the axis to extract a single slice
        extracted_slice = data[tuple(slices)]

        # Crop the 2D slice around the center in the remaining dimensions
        remaining_axes = [i for i in range(3) if i != axis]
        cropped_slice = extracted_slice[
            max(center[remaining_axes[0]] - half_size, 0):min(center[remaining_axes[0]] + half_size, extracted_slice.shape[0]),
            max(center[remaining_axes[1]] - half_size, 0):min(center[remaining_axes[1]] + half_size, extracted_slice.shape[1]),
        ]

        # Pad the slice to ensure 180x180 dimensions
        pad_height = slice_size - cropped_slice.shape[0]
        pad_width = slice_size - cropped_slice.shape[1]
        padded_slice = np.pad(cropped_slice, 
                              ((pad_height // 2, pad_height - pad_height // 2), 
                               (pad_width // 2, pad_width - pad_width // 2)),
                              mode='constant', constant_values=0)
        return padded_slice

    # Extract slices in axial, coronal, and sagittal planes
    axial_slice = extract_2d_slice(resampled_data, center, axis=2)  # Axial (z-axis)
    coronal_slice = extract_2d_slice(resampled_data, center, axis=1)  # Coronal (y-axis)
    sagittal_slice = extract_2d_slice(resampled_data, center, axis=0)  # Sagittal (x-axis)

    # Apply rotations to each slice
    axial_slice = np.rot90(axial_slice, k=-1)  # 90 degrees clockwise
    coronal_slice = np.rot90(coronal_slice, k=1)  # 90 degrees anticlockwise
    coronal_slice = np.rot90(coronal_slice, k=2)  # Additional 180 degrees
    sagittal_slice = np.rot90(sagittal_slice, k=1)  # 90 degrees anticlockwise
    sagittal_slice = np.rot90(sagittal_slice, k=2)  # Additional 180 degrees

    # Create subplots
    fig, axes = plt.subplots(1, 3, figsize=(12, 4))

    # Plot each padded and rotated slice
    axes[0].imshow(axial_slice, cmap="gray", origin="lower")
    axes[0].axis("off")

    axes[1].imshow(coronal_slice, cmap="gray", origin="lower")
    axes[1].axis("off")

    axes[2].imshow(sagittal_slice, cmap="gray", origin="lower")
    axes[2].axis("off")

    # Save the figure
    plt.tight_layout()
    plt.savefig(output_image_path, bbox_inches="tight", pad_inches=0)
    plt.close()
    
# Function to run nnUNet inference
@spaces.GPU(duration=70)  # Decorate the function to allocate GPU for its execution
def run_nnunet_predict(nifti_file):
    # Prepare directories
    os.makedirs(INPUT_DIR, exist_ok=True)
    os.makedirs(OUTPUT_DIR, exist_ok=True)

    # Extract the original filename without the extension
    original_filename = os.path.basename(nifti_file.name)
    base_filename = original_filename.replace(".nii.gz", "")
    
    # Save the uploaded file to the input directory
    input_path = os.path.join(INPUT_DIR, "image_0000.nii.gz")
    os.rename(nifti_file.name, input_path)  # Move the uploaded file to the expected input location
    
    # Debugging: List files in the /tmp/input directory
    print("Files in /tmp/input:")
    print(os.listdir(INPUT_DIR))

    # Set environment variables for nnUNet
    os.environ["nnUNet_results"] = MODEL_DIR

    # Construct and run the nnUNetv2_predict command
    command = [
        "nnUNetv2_predict",
        "-i", INPUT_DIR,
        "-o", OUTPUT_DIR,
        "-d", "004",                  # Dataset ID
        "-c", "3d_fullres",           # Configuration
        "-tr", "nnUNetTrainer_8000epochs",
        "-device", "cuda"  # Explicitly use GPU
    ]
    print("Files in /tmp/output:")
    print(os.listdir(OUTPUT_DIR))
    try:
        subprocess.run(command, check=True)

        # Rename the output file to match the original input filename
        output_file = os.path.join(OUTPUT_DIR, "image.nii.gz")
        new_output_file = os.path.join(OUTPUT_DIR, f"{base_filename}_LesionMask.nii.gz")
        if os.path.exists(output_file):
            os.rename(output_file, new_output_file)

            # Compute center of mass for the input image
            img = nib.load(input_path)
            data = img.get_fdata()
            affine = img.affine
            resampled_data, _ = resample_to_isotropic(data, affine, target_spacing=1.0)
            com = center_of_mass(resampled_data > 0)  # Center of mass
            center = np.round(com).astype(int)        # Round to integer

            # Extract and save 2D slices
            input_slice_path = os.path.join(OUTPUT_DIR, f"{base_filename}_input_slice.png")
            output_slice_path = os.path.join(OUTPUT_DIR, f"{base_filename}_output_slice.png")
            extract_middle_slices(input_path, input_slice_path, center=center)
            extract_middle_slices(new_output_file, output_slice_path, center=center)

            # Return paths for the Gradio interface
            return new_output_file, input_slice_path, output_slice_path
        else:
            return "Error: Output file not found."
    except subprocess.CalledProcessError as e:
        return f"Error: {e}"

# Gradio interface with adjusted layout
with gr.Blocks() as demo:
    gr.Markdown("""
   # 🔥 FLAMeS: FLAIR Lesion Segmentation for Multiple Sclerosis

    Upload a skull-stripped FLAIR brain MRI in NIfTI (.nii.gz) format to generate a binary segmentation of multiple sclerosis lesions.  
    FLAMeS is based on the nnUNet framework<sup>2</sup> and was trained on 668 MRI scans acquired using Siemens, GE, and Philips 1.5T and 3T scanners<sup>1</sup>.
    For skull-stripping, we suggest using [SynthStrip](https://surfer.nmr.mgh.harvard.edu/docs/synthstrip/) with the `--no-csf` flag for optimal results.

    Inference takes approximately 1 minute per MRI, with processing limited to one scan at a time due to Hugging Face's zero-GPU usage constraints. To process multiple cases simultaneously, download [FLAMeS's model](https://huggingface.co/FrancescoLR/FLAMeS-model) and run it locally using your own GPU or CPU setup.
    
    **Disclaimer:** Uploaded data is stored temporarily, no one has access to it, and it is deleted when the app is closed. For details, see [Gradio's file access guide](https://www.gradio.app/main/guides/file-access). Human subjects data should only be uploaded for processing if permitted by your institution's human subjects protection office.
    This is a research tool and is not intended for clinical use. Clinical decisions should not be based on the outputs of this tool.
    
    """)

    with gr.Row():
        with gr.Column(scale=1):
            flair_input = gr.File(label="Upload a FLAIR Image (.nii.gz)")
            submit_button = gr.Button("Submit")
        with gr.Column(scale=2):
            seg_output = gr.File(label="Download the Lesion Segmentation Mask")
            input_img = gr.Image(label="Input: FLAIR image")
            output_img = gr.Image(label="Output: Lesion Mask")

    gr.Markdown("""
    **If you find this tool useful, please consider citing:**

    1. A Deep Learning-Based Pipeline for Longitudinal White Matter Lesion Segmentation Using Diverse FLAIR Images  
   F. La Rosa, J. Dos Santos Silva, W. A. Mullins, H. Greenspan, J. F. Sumowski, D. S. Reich, & E. S. Beck.  
   *ACTRIMS Forum 2023. Multiple Sclerosis Journal.* 2023;29(2_suppl):18-242.  
   DOI: [10.1177/13524585231169437](https://doi.org/10.1177/13524585231169437)

    2. nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomedical Image Segmentation  
   F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, & K. H. Maier-Hein.  
   *Nature Methods.* 2021;18(2):203-211.  
   DOI: [10.1038/s41592-020-01008-z](https://www.nature.com/articles/s41592-020-01008-z)
    """)

    submit_button.click(
        fn=run_nnunet_predict,
        inputs=[flair_input],
        outputs=[seg_output, input_img, output_img]
    )

# Debugging GPU environment
if torch.cuda.is_available():
    print(f"GPU is available: {torch.cuda.get_device_name(0)}")
else:
    print("No GPU available. Falling back to CPU.")
    os.system("nvidia-smi")

download_model()

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