Self-Supervised Learning And Hybrid Deep Models For Predicting The Progression Of Fuchs' Endothelial Corneal Dystrophy To Endothelial Keratoplasty After Cataract Surgery

To evaluate the effectiveness of a self-supervised learning paradigm as a pre-training mechanism for hybrid deep learning models that integrate clinically relevant corneal biomarkers based on Scheimpflug maps into Convolutional Neural Networks (CNNs). This approach aims to predict the risk of progression of Fuchs' endothelial corneal dystrophy (FECD) to endothelial keratoplasty following cataract surgery. Our method seeks to improve prognostic accuracy compared to a previously published logistic regression model based on Scheimpflug tomography maps.

A multicenter prospective study analyzing Scheimpflug-based tomography images from tertiary ophthalmology centers, focusing on patients with FECD scheduled for cataract surgery.

We developed a hybrid model that incorporates clinically relevant corneal biomarkers—pachymetry, elevation and densitometry maps—through novel processing layers. The model was trained and validated on a dataset of 240 eyes of patients with vision loss, FECD, and cataracts, who underwent a cataract surgery. To overcome data limitations, we introduced a novel self-supervised learning task using an additional longitudinal dataset comprising 254 patients, each with multiple tomography images collected over several sessions. Model performance with and without pre-training through self-supervision was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, and specificity. 

The hybrid deep learning model pre-trained with self-supervision achieved an AUC of 0.9408, outperforming both the baseline model without pre-training (AUC = 0.9318) and the traditional logistic regression model (AUC = 0.9188). The proposed model demonstrated superior predictive ability for FECD progression to endothelial keratoplasty, significantly reducing false negatives compared to regression-based methods. Sensitivity and Specificity of this model was 90% and 85% respectively. This enhanced predictive capacity enables better risk stratification and more informed preoperative decision-making.

The combination of deep learning and clinically relevant corneal biomarkers markedly improves preoperative prognostic accuracy for FECD patients undergoing cataract surgery. Self-supervised learning proves to be a powerful approach when training data is limited. Our AI-driven model offers a valuable tool for identifying high-risk patients who may require endothelial keratoplasty, facilitating early surgical planning and optimizing visual outcomes. By improving sensitivity while maintaining clinically meaningful specificity, this approach supports better-informed surgical decisions in ophthalmology.