Deep learning-based diagnostic classification of multiple sclerosis using multicenter optical coherence tomography data

Abstract

Background

Multiple sclerosis (MS) is a chronic inflammatory disorder of the central nervous system, where timely and accurate diagnosis is essential for effective management. Optical coherence tomography (OCT) enables non-invasive evaluation of retinal changes that may serve as biomarkers for MS. Unlike other ophthalmologic diseases, raw cross-sectional OCT images in MS show subtle alterations often indistinguishable from healthy controls (HCs). Consequently, retinal layer thickness and boundary-derived surface features offer greater discriminatory power.

Methods

We investigated three categories of artificial intelligence (AI) models: (1) feature extraction with auto-encoder (AE) and shallow networks, (2) custom-designed deep networks, and (3) fine-tuned pre-trained networks. Retinal layer thickness and surface maps derived from OCT were analyzed to determine the most informative features, with channel-wise combination and mosaicing applied for feature integration. Model interpretability was assessed using occlusion sensitivity and Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations. The dataset included 38 HC and 78 MS eyes obtained from independent public and local sources. Patient-wise partitioning was implemented to prevent data leakage.

Results

The proposed deep network using channel-wise combined thickness maps of retinal nerve fiber layer (RNFL), ganglion cell and inner plexiform layer (GCIPL), and inner nuclear layer (INL) layers achieved balanced accuracy of 97.3% (SD = 4.16; 95% CI: 92.3–100%), specificity of 97.3% (SD = 5.59; 95% CI: 92.6–100%), sensitivity of 97.4% (SD = 3.54; 95% CI: 92.6–100%), g-mean of 97.3% (SD = 4.18; 95% CI: 92.24-100%), F1-score of 98.0% (SD = 3.86; 95% CI: 92.6–100%), and an AUC of 0.96 (SD = 0.08; 95% CI: 0.95–1.00). Notably, the high performance observed in internal cross-validation was achieved when public and local datasets were combined. However, performance decreased substantially in cross-dataset evaluations, where models were trained on one dataset and tested on the other, indicating limited external generalizability, particularly when trained on public data and applied to local clinical data.

Conclusions

AI-based analysis of OCT-derived retinal layer features enables accurate and interpretable classification of MS, supporting its potential as a valuable clinical biomarker.