Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Disc-Centric Contrastive Learning for Lumbar Spine Severity Grading

Sajjan Acharya, Pralisha Kansakar

TL;DR

This work tackles automated lumbar spinal stenosis severity grading from sagittal T2-weighted MRI by learning disc-centric representations within localized ROIs. A disc-centric contrastive pretraining framework is coupled with a coordinate-guided ROI localization auxiliary task and differential fine-tuning against a weighted focal loss objective, enabling robust disc-level feature learning and improved downstream grading. The approach achieves a balanced accuracy of $0.781$ ($78.1\%$) and reduces severe-to-normal misclassifications to $2.13\%$, outperforming training from scratch and a frozen encoder, while ROI localization attains $RMSE=10.18$ pixels. This disc-centric pipeline enhances safety and scalability for clinical deployment, though moderate-grade classification remains challenging and future work will explore multi-sequence MRI and ordinal modeling to further improve performance.

Abstract

This work examines a disc-centric approach for automated severity grading of lumbar spinal stenosis from sagittal T2-weighted MRI. The method combines contrastive pretraining with disc-level fine-tuning, using a single anatomically localized region of interest per intervertebral disc. Contrastive learning is employed to help the model focus on meaningful disc features and reduce sensitivity to irrelevant differences in image appearance. The framework includes an auxiliary regression task for disc localization and applies weighted focal loss to address class imbalance. Experiments demonstrate a 78.1% balanced accuracy and a reduced severe-to-normal misclassification rate of 2.13% compared with supervised training from scratch. Detecting discs with moderate severity can still be challenging, but focusing on disc-level features provides a practical way to assess the lumbar spinal stenosis.

Disc-Centric Contrastive Learning for Lumbar Spine Severity Grading

TL;DR

This work tackles automated lumbar spinal stenosis severity grading from sagittal T2-weighted MRI by learning disc-centric representations within localized ROIs. A disc-centric contrastive pretraining framework is coupled with a coordinate-guided ROI localization auxiliary task and differential fine-tuning against a weighted focal loss objective, enabling robust disc-level feature learning and improved downstream grading. The approach achieves a balanced accuracy of () and reduces severe-to-normal misclassifications to , outperforming training from scratch and a frozen encoder, while ROI localization attains pixels. This disc-centric pipeline enhances safety and scalability for clinical deployment, though moderate-grade classification remains challenging and future work will explore multi-sequence MRI and ordinal modeling to further improve performance.

Abstract

This work examines a disc-centric approach for automated severity grading of lumbar spinal stenosis from sagittal T2-weighted MRI. The method combines contrastive pretraining with disc-level fine-tuning, using a single anatomically localized region of interest per intervertebral disc. Contrastive learning is employed to help the model focus on meaningful disc features and reduce sensitivity to irrelevant differences in image appearance. The framework includes an auxiliary regression task for disc localization and applies weighted focal loss to address class imbalance. Experiments demonstrate a 78.1% balanced accuracy and a reduced severe-to-normal misclassification rate of 2.13% compared with supervised training from scratch. Detecting discs with moderate severity can still be challenging, but focusing on disc-level features provides a practical way to assess the lumbar spinal stenosis.
Paper Structure (45 sections, 5 equations, 4 figures, 1 table)

This paper contains 45 sections, 5 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Clinical Background
  3. Challenges in Automated MRI-Based Severity Classification
  4. Our Core Hypothesis
  5. Contributions
  6. Related Work
  7. Spine MRI Analysis
  8. Volumetric Models and Reconstruction
  9. Anatomical Localization
  10. Contrastive Learning in Medical Imaging
  11. Methodology
  12. Dataset and Preprocessing
  13. Data Background
  14. Data Composition
  15. Clinical Annotation
  16. ...and 30 more sections

Figures (4)

  • Figure 1: Visualization of disc localization and cropping. (a) Sagittal T2-weighted lumbar spine MRI slice showing the five intervertebral disc locations (L1/L2 through L5/S1) marked with colored dots. (b) Corresponding $96\times96$ pixel cropped Regions of Interest (ROIs) for each disc. Colors of the dots in (a) match the labels in (b).
  • Figure 2: Training dynamics of the contrastive representation learning framework. (a) Evolution of the multi-positive InfoNCE loss for training (blue) and validation (orange) sets over 60 epochs. (b) Learning rate trajectory utilizing a Cosine Annealing schedule.
  • Figure 3: Clinical performance and convergence analysis. (a) Validation balanced accuracy over 37 epochs. (b) Class-specific recall trajectories (Normal, Moderate, Severe); bold lines represent smoothed trends (EMA), dashed line at Epoch 21 marks the optimal checkpoint. (c) Confusion matrix at Epoch 21.
  • Figure 4: Visualization of model performance on a representative patient series (L1/L2, L2/L3). Each row corresponds to a vertebral level. Left: Original MRI with Ground Truth (green circle). Right: Model prediction (red cross) versus Ground Truth.