SAA: Scale-Aware Attention Block For Multi-Lesion Segmentation Of Fundus Images

Abstract

Multiple lesion segmentation, namely the segmentation of microaneurysms, soft exudate, hard exudate, and haemorrhage is very important to diabetic retinopathy diagnosis. However, the scales of different kinds of lesions are inconsistent. This inconsistent scale problem is unavoidable in the unified architecture design in which identical time of downsampling operations is used for different kinds of lesions. To achieve better performance at different scales, multiscale features need to be captured and adjusted. In this paper, we simply consider features from different stages of an encoder-decoder network as multiscale features. To re-weight importance of multiscale features dynamically, a scale-aware attention (SAA) block which consists of a spatial path and a channel path is introduced. In SAA block, adjusting operations are performed scale-wise instead of channel-wise or uniformly for all scales. Extensive experiments were conducted on two publicly-available datasets to verify the effect of SAA. SAA surpasses popular attention blocks and state-of-the-art results in the overall evaluation while comparable performance can be achieved in the individual evaluation at the same time.