Investigating the Impact of Semi-Supervised Learning Methods to Improve the Quality of Diagnosis of Retinal Diseases from OCT Images

Age-related Macular Degeneration (AMD) is a major cause of irreversible vision loss, especially in elderly, that can be diagnosed from OCT images. OCT, a noninvasive imaging modality, is not only widely applied in detecting retinal diseases but meaningful correlations have also been found between OCT images and neurological disorders like Alzheimer disease. The insufficient labeled dataset is the key challenge in using OCT images for disease detection, highlighting the importance of employing semi-supervised methods to address this issue. In this paper, the first step is to investigate the optimal structure of a supervised model for detecting AMD with the model being based on EfficientNet. In subsequent steps, the dataset size is reduced to 70%, 50%, 20%, and 5% of the total dataset to identify the best model based on an iterative teacher-student approach for detecting AMD disorders. In this study, the available OCT dataset gathered at Noor Eye Hospital consisting of 16,822 retinal OCT images are utilized. The optimized supervised model achieved 87.14% accuracy in distinguishing different AMD stages. As the dataset size is reduced to the most severe conditions (i.e., 20% and 5%), an expected decrease in accuracy to 77.05% and 54.78% has observed. Introduced semi-supervised learning based on iterative teacher-student model improved the accuracy to 88.56% at 20% and 64.15% at 5% volume of dataset, achieving high confidence levels, thereby enhancing the performance of the supervised model. This study introduces a framework that can be used in future studies to detect diseases with an insufficient OCT dataset, aiming to improve model accuracy.