Protovision: utilizing prototypical networks for retinal diseases classification based on few-shot learning

Abstract

Classifying retinal diseases with a higher accuracy rate is one of the most important means in the medical field. In the case of image classification, finding a dataset becomes a significant challenge for such cases. As a result, the accuracy rate of classification keeps deteriorating. To address this issue of data scarcity and improve the accuracy rate, the Few-Shot method has been proposed. The few-shot learning algorithms integrated into upgraded image classification techniques have been used to enhance retinal images. VGG19 and ResNet50 have been used for feature extraction and VGG19 has given promising results comparatively. Nonetheless, a variation of training episodes was evaluated to acquire the optimal outcome. The proposed method was tested on 4 new classes that are completely different from the training classes and 82% test accuracy was obtained. This acquired result leaves a further scope for potential applications of Few-Shot learning techniques in this medical field.