Deep Learning-based 2D segmentation and volume quantification of retinal detachments in OCT imaging of Vogt-Koyanagi-Harada disease

Mariana Dias Parrinha

Key information

Authors:

Mariana Dias Parrinha (Mariana Dias Parrinha)

Supervisors:

Diogo João Fróis Vieira (Diogo Fróis Vieira); João Miguel Raposo Sanches (João Miguel Raposo Sanches)

Published in

November 19, 2025

Abstract

Vogt-Koyanagi-Harada (VKH) is a rare autoimmune disorder accompanied by panuveitis, leading to choroidal inflammation, subretinal fluid (SRF) accumulation, and vision-threatening retinal detachments. Optical coherence tomography (OCT) is a key imaging tool for assessing ocular structures and tracking VKH progression. Among the most prominent OCT features of VKH are serous retinal detachments (SRDs), which serve as direct indicators of disease severity. Quantifying these detachments, particularly through volumetric assessment, enables more precise evaluation of disease activity and strengthens clinical prognostication. However, the distortion of retinal architecture caused by SRD, along with noise and low contrast in OCT images, makes SRD segmentation challenging. Moreover, the current literature provides limited segmentation studies for this rare disease and lacks publicly available code. Therefore, robust segmentation methods capable of handling VKH-specific complexity are required. This thesis proposes a U-Net model with attention mechanisms for SRD segmentation, achieving state-of-the-art performance, with mean DSC and IoU accuracies around 92% and 86%, respectively. Additionally, intra-slice alignment based on retinal layer detection is integrated to correct motion artifacts and ensure consistent volumetric reconstruction across follow-up scans. A 3D visualization module enables clinicians to assess the morphology and spatial extent of SRD with more clarity. Finally, several quantitative biomarkers are extracted, along with fluid volume metrics, providing a more comprehensive characterization of disease dynamics and supporting improved monitoring and prognostic evaluation in VKH patients.