Base on Megapixel Color Fundus Photos for Multi-label Disease Classification

Abstract

This paper discusses a new challenge of artificial intelligence in predicting fundus diseases: using only unprocessed million pixel Color Fundus Photos(CFP) to complete multi-label multi classification and lesion location tasks at the same time. In order to solve this problem, Double Flow Multi Instance Neural Network(DF-MINN) is designed. Df-MINN is an end-to-end dual flow network. It uses Multi Instance Spatial Attention(MISA) module to extract local information and Global Priorities Network base on Involvement(GPNI) module to analyze the overall content. In addition, experiments on the open multi label fundus dataset OIA-ODIR showed that DF-MINN was higher average precision than the previous network in the prediction of all seven diseases. Ablation experiments further proved the importance of high-resolution images in the diagnosis of fundus diseases.