Detecting Diverse Retinal Lesions Based on Low-Rank-Plus-Sparse Factorization Over Normal Retinal Images

摘要

This paper focuses on the detection of diverse lesions from retinal images. It is a challenging task because lesion number and lesion types in retinal images are generally unknown in advance, and different lesions may exhibit diverse visual properties. Many computer-aided detection methods have been developed for the automatic detection of specific lesions. However, they cannot well detect diverse lesions from retinal images. This paper develops a novel and effective framework to solve this issue. In the framework, instead of using lesion features or training images with labeled lesions, we directly learn the retinal background from normal retinal images for a test image and treat lesions as outliers of these normal images. To implement this idea, we collect a large number of normal retinal images and do some preprocessing steps on the normal set to acquire pure background images. Then a low-rank matrix is constructed by combined these backgrounds with the test image, and finally various lesions are well separated as sparse outliers by applying a low-rank-pluss-parse decomposition model to the matrix. The proposed method is evaluated on a dataset with diverse types of lesions. It obtains a high AUC value (0.9891) and a MAP value (0.8573), and is superior to the state-of-the-art methods, such as recently developed U-net, CNN, etc.