Three-dimensional OCT segmentation of anatomic and pathologic features of vitreoretinal interface disorders

We address the critical need for accurate and efficient segmentation of retinal structures in three-dimensional (3D) optical coherence tomography (OCT) scans, a key step in diagnosing and managing vision-threatening conditions. Leveraging the nnUNet architecture, we developed a semantic segmentation model that accurately identifies anatomical and pathological features in OCT images. To ensure high-quality training data, we employed a meticulous tiered labeling strategy. Initial annotations were performed using 3DSlicer by trained medical students, with subsequent review and refinement by senior and experienced medical students and ultimately reviewed by licensed ophthalmologists for complex cases. This meticulous approach ensured precise ground truth labels, enabling our model to achieve strong performance as measured by Dice similarity coefficients and pixel-level accuracy when compared to these ground truth labels in five-fold cross-validation. These metrics indicate the model’s ability to reliably delineate both normal anatomy and pathological structures, addressing limitations of prior methods that were often constrained to 2D datasets and small training set sizes due to labor-intensive labeling processes. Our work demonstrates the potential of advanced machine learning techniques to streamline OCT analysis, improving both the speed and reliability of retinal disease screening and diagnosis. By focusing on robust and generalizable segmentation, we aim to bridge the gap between the increasing demand for OCT interpretation and the limited availability of expert diagnosticians in screening environments. Furthermore, this foundation paves the way for future AI-assisted surgical and robotic systems that require precise real-time differentiation of retinal structures, ultimately improving outcomes for patients undergoing ophthalmic interventions.