OCT-SelfNet: A Self-Supervised Framework with Multi-Source Datasets for Generalized Retinal Disease Detection

Fatema-E Jannat^*Sina GholamiMinhaj Nur AlamHamed Tabkhi

• University of North Carolina at Charlotte, Charlotte, United States

In the medical AI field, there is a significant gap between advances in AI technology and the challenge of applying locally trained models to diverse patient populations. This is mainly due to the limited availability of labeled medical image data, driven by privacy concerns.To address this, we have developed a self-supervised machine learning framework for detecting eye diseases from optical coherence tomography (OCT) images, aiming to achieve generalized learning while minimizing the need for large labeled datasets.Our framework, OCT-SelfNet, effectively addresses the challenge of data scarcity by integrating diverse datasets from multiple sources, ensuring a comprehensive representation of eye diseases.By employing a robust two-phase training strategy self-supervised pre-training with unlabeled data followed by a supervised training stage, we utilized the power of a masked autoencoder built on the SwinV2 backbone.Extensive experiments were conducted across three datasets with varying encoder backbones, assessing scenarios including the absence of self-supervised pre-training, the absence of data fusion, low data availability, and unseen data to evaluate the efficacy of our methodology.OCT-SelfNet outperformed the baseline model (ResNet-50, ViT) in most cases. Additionally, when tested for cross-dataset generalization, OCT-SelfNet surpassed the performance of the baseline model, further demonstrating its strong generalization ability. An ablation study revealed significant improvements attributable to self-supervised pre-training and data fusion methodologies. Our findings suggest that the OCT-SelfNet framework is highly promising for real-world clinical deployment in detecting eye diseases from OCT images. This demonstrates the effectiveness of our two-phase training approach and the use of a masked autoencoder based on the SwinV2 backbone. Our work bridges the gap between basic research and clinical application, which significantly enhances the framework's domain adaptation and generalization capabilities in detecting eye diseases.