A Cross-Lesion Attention Network for Accurate Diabetic Retinopathy Grading With Fundus Images
Xiang Liu, Wei Chi
Abstract
Diabetic retinopathy (DR) is one of the major causes of blindness in the working-age population. Automatic DR grading with deep learning can help ophthalmologists treat patients in a timely manner. However, it is difficult to accurately grade DR because fundus images contain several different DR-related lesions such as soft exudates, hemorrhages, microaneurysms, and hard exudates, which are largely different in shape, appearance, and spatial location, and there exists strong multi-lesion dependency that can greatly affect the final grading results. In this paper, a Cross-Lesion Attention Network (CLANet) is proposed, which can adaptively learn rich and discriminative imaging features from complicated lesions and model the dependencies of DR-related lesions. It consists of two main parts as follows. First, an adaptive lesion-aware module based on adaptive channel-spatial convolution (ACSConv) can dynamically adjust convolution filters with different receptive fields to learn imaging features according to different DR lesions, so as to more flexibly and robustly handle the diversity of DR lesion features. Then, a cross-scale context attention module is developed to explore the dependencies of multiple DR-related lesions by sufficiently taking the advantage of the long-range dependence learning ability of context network, which is further aggregated to learn multi-scale context features in a coarse-to-fine manner. Experimental results on the public DR grading datasets (i.e., IDRiD, Messidor, and DDR) show that the proposed CLANet outperforms the state-of-the-art approaches.