Modular Transformer-Based Monitoring and Prediction of Wet AMD Progression Using OCT Imaging

Lovre Antonio Budimir, Ivana Zadro Matovinovic, Donik Vrsnak, Sven Loncaric

Abstract: Age-related macular degeneration (AMD) is a globally recognized age-related eye disease that can lead to severe visual impairment.This paper presents a novel application of modern computer vision techniques to monitor and predict the progression of wet AMD using Optical Coherence Tomography (OCT) imaging. The research in our work is structured around two tasks. The first task involves analyzing data from consecutive follow-up exams of a patient with wet AMD. In order to classify the evolution of neovascular activity between pairs of 2D OCT slices from these exams, a novel, fully transformer-based, three-part method is used to extract features from OCT B-scans and accurately detect and quantify changes in disease progression. The second task focuses on predicting neovascular activity within three months, using data from the current OCT exam. The goal is to develop a predictive model that can indicate the progression of wet AMD, enabling more informed and timely treatment planning. For the second task, we transfer knowledge from the key elements of the first task, which is considered simpler, into a two-part transformer-based method for extracting features and predicting disease evolution. We validate our methods on the MARIO-MICCAI 2024 Challenge dataset. The code is available at https://github.com/LovreAB17/FERLIV-MARIO.

External IDs:dblp:conf/miccai/BudimirMVL24