Pixel-Level Segmented Longitudinal Optical Coherence Tomography Data of Macular Hole Surgery Outcomes

Abstract

Research involving idiopathic full-thickness macular hole (iFTMH) has been limited by a shortage of longitudinal, annotated public datasets. In this article, we introduce a longitudinal annotated dataset containing OCT scans from patients who previously had iFTMH surgery with multiple time points. The OCT dataset based on the publicly accessible cohort contains a total of 2,591 fovea-centered horizontal and vertical B-scans that were collected at one preoperative and six postoperative time points (from 2 weeks to 48 months) that allow researchers to examine retinal recovery. We provide high-quality, expert-validated, pixel-level segmentation masks for 12 retinal features, representing both pathological (e.g., macular hole, cysts, epiretinal membrane) and anatomical (e.g., external limiting membrane, ellipsoid zone, retinal pigment epithelium) structures. Technical validation revealed high inter- and intra-rater reliability for each annotation (e.g., >0.80 average Dice coefficient for selected features). We also provide a reference nnUNet baseline to facilitate future benchmarking and method comparison. Overall, this resource is intended to facilitate AI modeling, support quantitative studies of retinal healing, and verify new ophthalmic image analysis methods.