This repository contains the code and methodology for a deep learning-based diagnostic system that assists in the detection and classification of lumbar spine degenerative conditions using MRI data. The project aims to reduce diagnostic time and support radiologists in clinical decision-making.
Low back pain, often caused by spondylosis and disc degeneration, affects over 619 million people globally. MRI is vital for assessing these conditions but is time-consuming and labor-intensive. This project proposes an AI-driven two-stage pipeline to detect and classify five common spinal disorders across lumbar disc levels (L1/L2 to L5/S1) with severity levels (Normal, Mild, Moderate, Severe).
-
Stage 1: Localization
- 3D ConvNeXt model to identify instance numbers (MRI slice levels)
- 2D CNN to predict precise (x, y) coordinates of vertebral disc levels
-
Stage 2: Severity Classification
- Uses image patches based on coordinates from Stage 1
- MIL (Multiple Instance Learning) + Bi-LSTM for contextual analysis
- Specialized preprocessing and augmentation for robustness
- Left/Right Neural Foraminal Narrowing
- Left/Right Subarticular Stenosis
- Spinal Canal Stenosis
- Initial experiments show the model accurately localizes disc levels
- Promising results from visual and quantitative evaluations
- Confusion matrices and performance graphs included
- Integrate severity scoring models
- Improve generalization with multi-task learning
- Finalize an end-to-end diagnostic pipeline ready for clinical use