RSNA 2024 Lumbar Spine Degenerative Classification

This project is designed to participate in the RSNA 2024 Lumbar Spine Degenerative Classification competition. The aim is to create machine learning models that simulate a radiologist's performance in diagnosing degenerative spine conditions using lumbar spine MRI images. The project focuses on multi-task classification of five key conditions related to lumbar spine degeneration, helping radiologists detect, assess severity, and classify conditions accurately.

Table of Contents

• Overview
• Dataset
• Model Architecture
• Installation
• Usage
• Docker Usage
• Preprocessing
• Training
• Evaluation
• Results
• Configuration
• License

Overview

Low back pain is the leading cause of disability globally, often attributed to degenerative conditions of the lumbar spine. Using MRI scans, the competition aims to classify conditions such as:

• Left Neural Foraminal Narrowing
• Right Neural Foraminal Narrowing
• Left Subarticular Stenosis
• Right Subarticular Stenosis
• Spinal Canal Stenosis

Participants develop models to classify the severity of these conditions across five vertebrae levels (L1/L2 to L5/S1) into three severity grades:

1. Normal/Mild
2. Moderate
3. Severe

Dataset

The dataset consists of MRI images in DICOM format, providing detailed medical imaging for the lumbar spine. Labels include severity scores for the above conditions. The dataset is multi-institutional, sourced from eight different sites across five continents.

Model Architecture

This project supports several advanced models for 3D MRI analysis using PyTorch and MONAI:

• ResNet3D: For volumetric feature extraction.
• SwinUNETR: A transformer-based architecture for medical imaging.
• SegResNet: A specialized segmentation model for volumetric data.
• DynUNet: A dynamic U-Net with adjustable depth.
• ViT: Vision Transformer for efficient spatial attention.

Multi-Task Learning:

The model performs multi-task classification on three labels (central, left, and right severity scores), using a combination of image data and metadata (e.g., slice thickness, pixel spacing).

Installation

Follow these steps to set up and run the project:

1. Clone the repository:

   git clone https://github.com/anto18671/lumbar-spine-degenerative-classification.git
   cd lumbar-spine-degenerative-classification

2. Install dependencies: Use the provided requirements.txt to install the necessary packages:

   pip install -r requirements.txt

3. Set up data: Place the dataset in the data/ directory.

Usage

To train and evaluate the model, run the following:

python main.py

Docker Usage

If you prefer using Docker to set up the environment and avoid dependency issues, follow these steps to run the project in a Docker container:

Using Prebuilt Docker Image from GitHub Container Registry:

Alternatively, you can pull the prebuilt image from GitHub Container Registry and run it without building the image locally.

docker pull ghcr.io/anto18671/lumbar-spine-degenerative-classification:latest

Then, run the image:

docker run --gpus all -v /path/to/data:/app/data ghcr.io/anto18671/lumbar-spine-degenerative-classification:latest

This approach ensures that your development environment is consistent and isolated from any external dependencies.

Key Features:

• Mixed Precision Training: Utilizes PyTorch's autocast for improved memory efficiency.
• Stratified Group K-Fold: Data split to preserve label distribution and prevent data leakage.

Preprocessing

Preprocessing includes:

• DICOM Loading: Loading images from DICOM format, correcting orientation, and handling pixel values.
• Spatial Scaling: Rescaling images based on pixel spacing for consistent physical dimensions.
• Augmentations: Elastic transformations, rotations, horizontal flips, and Gaussian noise applied during training.

Training

The training process uses a multi-task loss function, where each task (central, left, and right) is trained separately using class-weighted cross-entropy.

Training Configurations:

• Model architecture: ResNet3D, SwinUNETR, SegResNet, DynUNet, or ViT.
• Loss function: Multi-task cross-entropy with class weights.
• Optimizer: AdamW with weight decay and learning rate scheduling.
• Scheduler: Exponential learning rate decay.

Evaluation

After training, the model is evaluated on validation data using the following metrics:

• AUC-ROC: Evaluates the Area Under the Curve for Receiver Operating Characteristics across severity classes.
• Confusion Matrix: To visualize prediction accuracy per class.
• Classification Report: Includes precision, recall, and F1 scores for each severity class.

Results

The performance is measured by weighted log loss with the following sample weights:

• Normal/Mild: 1
• Moderate: 2
• Severe: 4

Configuration

Key configurations are defined in the CONFIG dictionary at the beginning of the script:

• enable_spatial_scaling: Enable or disable spatial scaling using pixel spacing.
• model_architecture: Choose between 'ResNet3D', 'SwinUNETR', 'SegResNet', 'DynUNet', or 'ViT'.

License

This project is licensed under the MIT License - see the LICENSE file for details.