Automatic detection and prediction of nAMD activity change in retinal OCT using Siamese networks and Wasserstein Distance for ordinality

Taha Emre 1∗, Teresa Araújo 2∗, Marzieh Oghbaie 2∗ Dmitrii Lachinov, Guilherme Aresta, Hrvoje Bogunović · 2025

Read Automatic detection and prediction of nAMD activity change in retinal OCT using Siamese networks and Wasserstein Distance for ordinality by Taha Emre 1∗, Teresa Araújo 2∗, Marzieh Oghbaie 2∗ Dmitrii Lachinov, Guilherme Aresta, Hrvoje Bogunović (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-14323.txt.

Introduction

Taha Emre, Teresa Araujo, Marzieh Oghbaie, and co-authors build deep learning tools for an eye disease in a 2025 paper. The disease, a wet form of macular degeneration, drives vision loss in older adults. Spotting change in activity early guides timely drug doses and better outcomes.

One model pairs scans from two visits in a twin network to call if severity changed. The other model forecasts change three months out, a first for this task. Its loss treats the severity grades as ordered steps, using a distance built for ranked classes. Both models ranked high on the board of a public contest held in 2024. The team sees clear use in planning care for this disease.

Catalog record

CREATORTaha Emre 1∗, Teresa Araújo 2∗, Marzieh Oghbaie 2∗ Dmitrii Lachinov, Guilherme Aresta, Hrvoje Bogunović
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-14323

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Automatic detection and prediction of nAMD activity change in retinal OCT using Siamese networks and Wasserstein Distance for ordinality Taha Emre 1∗, Teresa Araújo 2∗, Marzieh Oghbaie 2∗ Dmitrii Lachinov, Guilherme Aresta, Hrvoje Bogunović Abstract Neovascular age-related macular degeneration (nAMD) is a leading cause of vision loss among older adults, where disease activity detection and progression prediction are critical for nAMD management in terms of timely drug administration and improving patient outcomes. Recent advancements in deep learning offer a promising solution for predicting changes in AMD from optical coherence tomography (OCT) retinal volumes. In this work, we proposed deep learning models for the two tasks of the public MARIO Challenge at MICCAI 2024, designed to detect and forecast changes in nAMD severity with longitudinal retinal OCT. For the first task, we employ a Vision Transformer (ViT) based Siamese Network to detect changes in AMD severity by comparing scan embeddings of a patient from different time points. To train a model to forecast the change after 3 months, we exploit, for the first time, an Earth Mover (Wasserstein) Distance-based loss to harness the ordinal relation within the severity change classes. Both models ranked high on the preliminary leaderboard, demonstrating that their predictive capabilities could facilitate nAMD treatment management. ∗∗ ** ** footnotetext: https://github.com/EmreTaha/Siamese-EMD-for-AMD-Change * * footnotetext: These authors contributed equally to this work 1 Introduction Neovascular age-related macular degeneration (nAMD) is a progressive exudative disease characterized by the accumulation of fluid in the macula, which can significantly impair vision function [ 6 ] . Anti-VEGF treatments have shown great efficacy in mitigating AMD progression, and the positive effects are optimized by reducing the time from fluid onset to treatment, and thus regular follow-up is crucial for successful patient outcomes. The presence and extent of exudative signs, such as intraretinal and subretinal fluid, which are best visible on optical coherence tomography (OCT) images, are relevant markers for anti-VEFG administration [ 19 ] . Thus, predicting and accurately detecting changes in neovascularization activity are pivotal tasks for treatment management. The "Monitoring Age-related Macular Degeneration Progression In Optical Coherence Tomography" (MARIO) Challenge, organized as part of MICCAI 2024 [ 1 ] , aims the development of automatic methods for nAMD progression assessment and evolution prediction. 1.1 Longitudinal change detection in retinal OCT The manual assessment of changes between image pairs is time-consuming and challenging, highlighting the need for automated systems to detect meaningful changes and reduce specialists’ workload. Although machine learning, particularly deep learning, has advanced in automating OCT disease diagnostics, there has been little focus on detecting changes between sequential OCT images. In medical image analysis, current deep learning methods for change detection usually fall into two categories: 1) Siamese networks [ 13 , 12 , 21 ] , or 2) Graph-based methods [ 9 , 18 ] . In Li et al. [ 13 ] , change detection is approached as a metric-learning problem, using a Siamese neural network to assess changes between two patient visits. A contrastive loss function is applied between features of two images, labeled as change or no change , to output a pairwise distance between images from two time points. This method was tested on retinal photographs (diabetic retinopathy of prematurity) and knee x-rays (osteoarthritis). In Li et al. [ 12 ] , the same authors used a similar approach to track COVID-19 pulmonary disease severity in chest x-rays. In contrast, Karward et al. [ 9 ] proposed a graph-based, anatomy-aware model for tracking changes between chest X-rays by using both local and global anatomical information. This model provides localized comparisons between sequential X-ray images and outperformed Siamese-based models. However, this type of approach is more suitable when the structures are within a rigid region, and thus have a stable anatomical location (e.g., chest x-rays), and are not that suitable for retinal OCT, where pathologies can cause significant tissue deformations. 1.2 Prediction of nAMD evolution within 3 months on OCT nAMD severity change prediction is crucial for patient follow-up scheduling and timely drug (anti-VEGF) administration. In this aspect, it is essential to have a predictive model that can assess the change in AMD disease activity within a meaningful time-frame from an OCT scan acquired at a visit. Prior work has largely focused on a related task of intermediate AMD progression prediction [ 4 , 2 , 17 , 22 ] . The main challenge of training a predictive model of disease progression and treatment response is the training data. Since, the nature of the task is temporal, it is crucial to have follow-up visits from a patient to create a longitudinal dataset. In [ 4 , 2 , 17 ] , they used longitudinal self-supervised learning methods to learn the temporal relations. Additionally, most of the available OCTs have no observable change within short time windows, resulting in severe class imbalance, requiring specialized deep learning loss terms to address this [ 14 ] . 1.3 Contributions We contribute to the state of the art in longitudinal retinal OCT assessment in two ways. First, we propose a Siamese-based approach, SiamRETFound, relying on an OCT foundation model (RETFound [ 23 ] ) and additional pretraining for learning to estimate clinically-relevant changes between two patient B-scans. Second, we propose a classifier model that predicts the nAMD evolution within 3 months. Specifically, our novel loss has 2 parts: a focal loss to address the severe class imbalance, and an Earth Mover’s Distance-based (EMD) loss to harness the ordinal relation within the severity change classes. To our knowledge, this is the first study to use EMD for predicting disease severity evolution. 1.4 MARIO Challenge We assess our methods on the two tasks of the MARIO challenge. Task 1 (T1) aims at developing algorithms to classify changes in disease activity from pairs of retinal OCT B-scans from two visits of patients with nAMD undergoing anti-VEGF treatment, to support decision-making. The classification categories were Reduced, Stable, Worsened , or Other (Uninterpretable). Based on our qualitative assessment of the provided training set, Other seemed to be associated with scenarios that prevented proper clinical assessment, such as the presence of noise, obscured regions, vertical flipping of the scan, poor alignment between image pairs, etc. (Fig. 2 : (e)-(g)). However, no objective criteria was provided, and there were a few nuanced cases (Fig. 2 : (g)). In Task 2 (T2), the goal is to train a predictive model that can accurately predict the change in disease activity within 3 months, given a single B-scan. The prediction categories are Reduced , Stable and Worsened .Even though the labels and the prediction are on B-scan level, labels are consistent within a volume. 2 Method 2.1 T1: SiamRETFound for longitudinal change detection We propose SiamRETFound, a Siamese neural network with shared weights for evaluating change between retinal B-scans from two visits of the same patient (Fig. 1 ). SiamRETFound uses a late fusion approach by extracting and concatenating two feature representations, one for each B-scan. The resulting feature vector, which integrates meaningful features from both time points, is then processed by a classification head to obtain the change detection. Ultimately, SiamRETFound learns to assess which differences between the extracted features are relevant to identify clinically meaningful changes. The backbone for

Read the complete text