🧠 Lumbar Spine MRI Classification Using Deep Learning

📌 Project Overview

This project tackles the RSNA 2024 Lumbar Spine Degenerative Classification Challenge, which aims to classify the severity of five degenerative spine conditions using lumbar spine MRI scans. The work leverages deep learning techniques to automate grading of spinal stenosis and related pathologies, improving diagnostic efficiency and reducing variability in radiology assessments.

🎯 Objectives

• Develop a deep learning model to classify degenerative spine conditions.
• Preprocess MRI DICOM images into structured datasets suitable for training.
• Compare 2D CNN approaches on sagittal and axial slices.
• Evaluate performance on severity classification tasks (normal, mild, moderate, severe).

🛠️ Methods

• Data Preprocessing

  • Filtered MRI DICOMs for T2 sagittal and axial series.
  • Generated slice-based datasets with metadata (study ID, level, condition, severity).
  • Applied resizing, normalization, and mini-stack approaches.

• Modeling

  • Implemented a 2D CNN classifier in PyTorch.
  • Designed experiments for sagittal vs axial views across conditions.
  • Used cross-entropy loss and Adam optimizer for training.

• Evaluation

  • Metrics: Accuracy, F1-score, AUROC.
  • Compared results across different anatomical views and conditions.

📊 Key Findings

• Sagittal views were more effective for neural foraminal stenosis, while axial views worked better for subarticular and canal stenosis.
• The CNN achieved reasonable accuracy in classifying severity levels, though performance varied by condition.
• Identified opportunities to expand toward 3D CNNs, transfer learning, and multimodal ensembles for future work.

🚀 Future Directions

• Incorporate 3D volumetric modeling for better anatomical context.
• Leverage transformer-based architectures (Vision Transformers).
• Improve data augmentation and class balancing strategies.
• Conduct external validation on independent MRI datasets.

🧰 Tech Stack

• Python, PyTorch
• NumPy, pandas, matplotlib
• pydicom for medical imaging preprocessing
• scikit-learn for evaluation metrics

📂 Repository Structure

├── data/               # Preprocessed MRI slices / metadata  
├── notebooks/          # Jupyter notebooks for EDA & prototyping  
├── models/             # CNN architecture and training scripts  
├── results/            # Model outputs and evaluation metrics  
└── README.md           # Project documentation