Computer methods and programs in biomedicine · 2025
Self-supervised learning and hybrid deep models for predicting the progression of Fuchs' endothelial corneal dystrophy after cataract surgery.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
León-Tarife Paola, Arnalich-Montiel Francisco, Mingo-Botín David, Díaz de María Fernando, Stokking Martha, Saenz-Madrazo Nerea, Etxebarria-Ecenarro Jaime, Arriola-Villalobos Pedro, Boto de Los Bueis Ana, González-Díaz Iván
Аннотация
BACKGROUND AND OBJECTIVE: Fuchs' endothelial corneal dystrophy (FECD) increases the risk of corneal decompensation after cataract surgery, often leading to endothelial keratoplasty (EK). Reliable prediction of EK is essential for surgical planning, yet traditional corneal biomarkers have limited prognostic value.
METHODS: We present a novel deep learning framework based on Scheimpflug tomographic imaging to improve FECD prognosis. The approach integrates three components to address data limitations: (1) clinical domain knowledge, (2) ensemble learning, and (3) self-supervised learning (SSL). A hybrid convolutional neural network (CNN) is proposed, combining a RANSAC-based algorithm for estimating best-fit sphere (BFS) elevation maps with a dual-branch design that incorporates Polar Pooling to mimic clinical reasoning. Robustness is enhanced through bootstrap aggregation, where multiple models are trained on different data subsets and their predictions averaged. To further reduce reliance on annotated data, we introduce a self-supervised contrastive pretraining task. This task distinguishes images from the same eye across time, using a weighted contrastive loss that emphasizes corneal irregularities.
RESULTS: On a multi-center dataset, the framework achieved an AUC of 0.94, outperforming biomarker-based models and demonstrating strong generalizability across hospitals.
CONCLUSIONS: The proposed methodology provides a robust and interpretable decision-support tool for FECD management and cataract surgery planning.
METHODS: We present a novel deep learning framework based on Scheimpflug tomographic imaging to improve FECD prognosis. The approach integrates three components to address data limitations: (1) clinical domain knowledge, (2) ensemble learning, and (3) self-supervised learning (SSL). A hybrid convolutional neural network (CNN) is proposed, combining a RANSAC-based algorithm for estimating best-fit sphere (BFS) elevation maps with a dual-branch design that incorporates Polar Pooling to mimic clinical reasoning. Robustness is enhanced through bootstrap aggregation, where multiple models are trained on different data subsets and their predictions averaged. To further reduce reliance on annotated data, we introduce a self-supervised contrastive pretraining task. This task distinguishes images from the same eye across time, using a weighted contrastive loss that emphasizes corneal irregularities.
RESULTS: On a multi-center dataset, the framework achieved an AUC of 0.94, outperforming biomarker-based models and demonstrating strong generalizability across hospitals.
CONCLUSIONS: The proposed methodology provides a robust and interpretable decision-support tool for FECD management and cataract surgery planning.