MEDIA-Net: Multi-scale Feature Extraction and DepthwIse Attention Network for Accurate Carotid Artery Segmentation

Abstract

Accurate segmentation of the carotid artery in digital subtraction angiography (DSA) imaging is crucial for the diagnosis, prognosis and therapy planning of chronic carotid artery occlusion (CCAO). However, general segmentation methods often produce misclassifications and fuzzy boundaries, making it challenging to obtain precise segmentation results. This paper suggests a solution to these problems by proposing a Multi-scale Feature Extraction and DepthwIse Attention Network (MEDIA-Net) consisting of the Multi-scale Feature Extraction (MFE) module and the DepthwIse Attention (DA) module. The MFE module extracts and blends features at different scales to capture dependencies between multi-scale components. In contrast, the DA module further utilizes depthwise convolution, gated attention and spatial attention to enhance the network's classification and boundary delineation accuracy. Experimental results show that MEDIA-Net outperforms general methods regarding mIoU and Dice score, producing fewer misclassifications and s harper boundaries. The proposed method offers a promising solution for carotid artery segmentation in DSA imaging. It highlights the importance of densely combining multi-scale feature extraction and depthwise attention in general medical image segmentation.