Reducing Manual Labeling Requirements and Improved Retinal Ganglion Cell Identification in 3D AO-OCT Volumes Using Semi-Supervised Learning

Adaptive-optics OCT can resolve individual retinal ganglion cells in vivo, but building supervised models to identify them requires expert annotation of 3D volumes — slow, costly, and a bottleneck on study size. This work applies semi-supervised learning so that abundant unlabeled AO-OCT volumes contribute to training alongside a small labeled set. The approach reduces the manual labeling requirement substantially while improving retinal ganglion cell identification accuracy, making larger-scale structural studies of the ganglion cell layer practical.