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

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Multiple sclerosis (MS) is a central nervous system inflammatory disorder requiring precise diagnosis.
• Optical coherence tomography (OCT) detects retinal changes, potential MS biomarkers.
• Subtle OCT alterations in MS necessitate advanced analytical methods beyond raw image inspection.

Purpose of the Study:

• To evaluate artificial intelligence (AI) models for classifying multiple sclerosis (MS) using optical coherence tomography (OCT) derived retinal features.
• To determine the most informative retinal layer thickness and surface features for MS detection.
• To assess the interpretability and generalizability of AI models in MS classification.

Main Methods:

• Investigated three AI model categories: feature extraction with auto-encoders (AE) and shallow networks, custom deep networks, and fine-tuned pre-trained networks.
• Analyzed retinal layer thickness and surface maps from OCT, integrating features via channel-wise combination and mosaicing.
• Utilized occlusion sensitivity and Grad-CAM for model interpretability on a dataset of 38 healthy control (HC) and 78 MS eyes.

Main Results:

• A deep network combining retinal nerve fiber layer (RNFL), ganglion cell and inner plexiform layer (GCIPL), and inner nuclear layer (INL) thickness maps achieved 97.3% balanced accuracy.
• High performance was observed when combining public and local datasets for internal cross-validation.
• Performance significantly decreased in cross-dataset evaluations, highlighting limited external generalizability, especially when training on public data.

Conclusions:

• AI-based analysis of OCT-derived retinal features provides accurate and interpretable MS classification.
• This approach supports the potential of OCT-derived retinal biomarkers for MS diagnosis.
• Further research is needed to improve the generalizability of AI models across different datasets.