Classification of Myopic Maculopathy Images with Self-supervised Driven Multiple Instance Learning Network

Jiawen Li, Jaehyeon Soon, Qilai Zhang, Qifan Zhang, Yonghong He

Abstract: Myopia is a high-incidence disease that widely exists across various regions. If left unaddressed, it may escalate into high myopia. The leading cause of visual impairment is myopic maculopathy. Currently, certain deep-learning techniques have been employed for the analysis of images depicting myopic maculopathy in fundus photography. These methods are dedicated to assisting physicians in efficient disease diagnosis. In our work, a deep learning framework is introduced to classify images of five different severities of myopic maculopathy. First, we employ a diffusion model to generate a series of images for data augmentation to alleviate the pressure of uneven distribution of categories in training datasets, then we divide images into multiple patches and perform self-supervised learning to generate patch-level feature embeddings. Building upon the above foundation, an aggregator is proposed based on multiple instance learning to achieve image-level classification. We demonstrate the effectiveness of this method in four sufficient experiments with three key evaluation metrics of quadratic-weighted kappa, F1 score, and specificity. Our approach secured the tenth position in the Myopic Maculopathy Analysis Challenge 2023 (MICCAI MMAC 2023).