Machine-learning-based automated Schlemms canal volumetric segmentation for optical coherence tomography

Abstract

Volumetric segmentation of Schlemms Canal (SC) in optical coherence tomography (OCT) is time-consuming, creating a barrier to experiments studying glaucoma and the anatomy of the trabecular outflow pathways in vivo. To this end, we developed an automated segmentation tool, Schlemms Canal-Localization and Semantic Segmentation (SC-LSS), for the volumetric segmentation of SC in in vivo mice eyes from visible-light OCT (vis-OCT). SC-LSS first localizes the boundaries of SC and subsequently determines the boundaries of SC within the localized region. We used 324 B-scans from 16 mouse eyes for training, validation, and testing the model, and 203 additional B-scans from 16 mouse eyes to evaluate the models accuracy. We found that the Dice coefficient between segmentations generated by SC-LSS and manual expert graders was 0.70 {+/-} 0.20 and that the Dice coefficient between two expert graders was 0.73 {+/-} 0.18 (p = 0.10). Furthermore, SC-LSS captured decreases in SC size with increasing intraocular pressure, yielding a 51.5% decrease in SC size at 20 mmHg compared to 5 mmHg. SC-LSS also identified a 20.1% increase in SC size following the administration of pilocarpine. We anticipate that SC-LSS will accelerate studies on factors regulating the trabecular outflow pathways and their role in glaucoma development and management.