Improving Generalization of Multi-label Classification of Fundus Images with Adaptive Histogram Equalization

Abstract

Since reading fundus images is a task that requires a high degree of expertise and concentration, computer-aided diagnosis (CAD) can be an effective aid. In particular, the Concurrent Reader method, which predicts the names of findings corresponding to fundus images and provides them to the physician at the time of diagnosis, is precious from a medical-economic point of view because it can significantly reduce the time required for reading. In this study, we focus on a multi-label classification model that predicts the names of findings from fundus images, implemented in Concurrent Reader-based fundus reading CAD. We show that adaptive histogram equalization can improve the generalization performance of the multi-label classification model by eliminating differences among camera models. Furthermore, we evaluate and discuss the impact of the choice of parameters and color space when applying adaptive histogram equalization.