Scholarlink

Age-related Macular Degeneration (AMD) is a leading cause of legal blindness worldwide, with late-stage forms requiring close monitoring for effective treatment. Currently, this monitoring is performed manually by clinicians using Optical Coherence Tomography (OCT), a process that is both labor-intensive and time-consuming. Computer-assisted diagnosis models provide a scalable alternative. In this work, we propose a method for the MARIO challenge based on an Ensemble of Siamese networks to analyze two consecutive OCT B-scans and classify AMD progression. Our approach combines three state-of-the-art CNNs and incorporates categorical patient data and a 2D infrared OCT image within a customized fusion module. Additionally, we use Knowledge Distillation to train a neural network to predict disease evolution from a single B-scan, addressing the need for predicting future progression from one time point. Experimental results show the validity of our method for the classification task, achieving an F1-score of 0.832 on an external evaluation set.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-modal Siamese Ensemble for Neovascular AMD Classification and Prediction from Optical Coherence Tomography

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sebastien Richard,
  • Marie Beurton-Aimar

摘要

Age-related Macular Degeneration (AMD) is a leading cause of legal blindness worldwide, with late-stage forms requiring close monitoring for effective treatment. Currently, this monitoring is performed manually by clinicians using Optical Coherence Tomography (OCT), a process that is both labor-intensive and time-consuming. Computer-assisted diagnosis models provide a scalable alternative. In this work, we propose a method for the MARIO challenge based on an Ensemble of Siamese networks to analyze two consecutive OCT B-scans and classify AMD progression. Our approach combines three state-of-the-art CNNs and incorporates categorical patient data and a 2D infrared OCT image within a customized fusion module. Additionally, we use Knowledge Distillation to train a neural network to predict disease evolution from a single B-scan, addressing the need for predicting future progression from one time point. Experimental results show the validity of our method for the classification task, achieving an F1-score of 0.832 on an external evaluation set.