Work

OCT-Derived Radiomic Features for Differentiating Optic Neuritis Etiologies

Year: 2026

Type: article

Abstract: Purpose To assess the feasibility of an optical coherence tomography (OCT)-based radiomics approach in optic neuritis (ON) and determine whether peripapillary retinal nerve fiber layer (pRNFL) OCT-derived radiomic features could help differentiate between multiple sclerosis–associated optic neuritis (MS-ON) and myelin oligodendrocyte glycoprotein antibody–associated disease optic neuritis (MOGAD-ON). Design Retrospective cohort study. Participants Patients with ON associated with either multiple sclerosis (MS-ON) or MOGAD-ON were selected from the existing ONET (Optic Neuropathies of la Timone) cohort. Forty-five patients (61 eyes with optic neuritis; 28 MS and 17 MOGAD) contributing 266 OCT examinations were included. Methods A radiomics pipeline was applied to pRNFL OCT data to extract quantitative imaging features. A total of 102 radiomic features were generated and reduced using feature selection methods. Unsupervised and supervised machine-learning approaches were used to identify relevant radiomic patterns and evaluate their ability to discriminate MS-ON from MOGAD-ON. Models were trained using radiomic features alone, clinical variables alone, and a combination of both. Main Outcome Measures The primary outcome measure was the diagnostic performance of machine-learning models in distinguishing MS-ON from MOGAD-ON, assessed using the area under the receiver operating characteristic curve (AUC). Results Among the 102 extracted radiomic features, 62 were retained after feature reduction, and three features were selected through machine-learning–based analysis. Supervised machine learning models (Logistic Regression) using radiomic features outperformed models based on clinical variables alone (AUC 0.763 vs 0.584, p = 0.001). Combining radiomics with clinical variables yielded a numerically higher AUC of 0.788 compared with radiomics alone (0.763), but without statistical significancy (p = 0.361). The combined model remained significantly better than the clinical-variable-only model (p < 0.001). Conclusions This study establishes a reproducible OCT radiomics pipeline and demonstrates the feasibility of applying unsupervised and supervised machine-learning approaches to pRNFL OCT data. Radiomic features show promising potential for improving the differentiation between MOGAD-ON and MS-ON.

Source: Ophthalmology Science

DOI: 10.1016/j.xops.2026.101401