SMANet: Superpixel-guided multi-scale attention network for medical image segmentation

Abstract

Medical image segmentation plays a crucial role in assisting diagnosis. However, the inherent low contrast and noise in medical images make it challenging to achieve accurate medical image segmentation. To address this problem, we propose a superpixel-guided multi-scale attention network (SMANet) for segmenting medical images accurately. Superpixel segmentation could effectively divide medical images into different regions based on image gradient information. Accordingly, a superpixel-guided fusion attention module is proposed to utilize the regional division information provided by superpixel segmentation and further optimize the features in spatial and channel dimensions. In the encoder stage, an inverted pyramid feature extraction architecture is constructed to take advantage of texture information in shallow features, effectively solving the problem of information loss caused by sampling. In the proposed multi-scale feature joint decoder, multi-scale features are effectively enhanced and integrated to reconstruct image details guided by high-level features. Specifically, the full-scale feature attention module is embedded into multi-scale skip connections to contribute to the sufficient expression of important semantic information in features. Besides, we redesign the classic decoder to make full use of to the semantic information of deep features to guide feature fusion. Extensive experiments based on different public datasets and proposed neck vessels ultrasound dataset (USdata) prove the superiority of SMANet in terms of generalization, qualitative and quantitative performance.