Retinal fluid segmentation in real-world OCT imaging of neovascular AMD using hybrid deep learning and graph-based optimization

Abstract

Accurate segmentation and quantification of retinal fluid in optical coherence tomography (OCT) images are important for assessing disease activity and treatment response in neovascular age-related macular degeneration (nvAMD). Yet, manual delineation of intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED) is labor-intensive and subject to inter-grader variability. We developed a hybrid framework for robust retinal fluid segmentation in clinical-grade longitudinal OCT, combining anatomically constrained retinal-region extraction using Deep LOGISMOS with nnU-Net–based 2-D and 3-D fluid segmentation. To improve adaptation to heterogeneous clinical data while reducing annotation burden, an iterative expert-guided labeling strategy was used, starting with 70 RETOUCH OCT scans and 100 scans from subjects with macular neovascularization at the University of Iowa and expanding to a final aggregated training set of 291 scans. The method was validated on a held-out test set of 50 OCT volumes from 50 nvAMD subjects, with a subset of 20 volumes independently annotated by a second expert for inter-observer analysis. It was further applied to a longitudinal dataset of 14,500 Heidelberg Spectralis macular OCT scans from 221 independent subjects. The ensemble model achieved Dice similarity coefficients (%) of 89.5±11.7 for IRF, 88.0±10.7 for SRF, and 82.0±13.0 for PED against the primary expert annotations, with performance approaching expert-level agreement. In longitudinal inference, our developed framework supports robust large-scale quantification of fluid burden and spatial extent across heterogeneous scan protocols and extended follow-up. The artificial intelligence-based approach has the potential to enable scalable and clinically meaningful longitudinal analysis of retinal fluid in nvAMD.