Multiple Self-attention Network for Intracranial Vessel Segmentation

The capture of long-distance dependencies presents an efficient approach for feature learning and extraction. Especially the Transformers models that explore the dependencies within a long sequence, have swept the fields of natural language processing with their powerful performance. However, Transformers require extremely high computing power due to the huge amount of parameters, and it cannot achieve parallelism since it outputs tokens one by one. In this work, inspired by Transformers, we propose a self-attention encoder module (SAEM) that focuses on learning the connections between each position and all other positions in the image, which preserves the efficiency of Transformers but with less calculation and faster inference speed. In our SAEM, different group of internal feature maps within images captured by multiple scaled self-attentions are cascaded to generate global context information. Based on our SAEM, a lightweight and parallel network is designed for segmentation of intracranial blood vessels. Moreover, a data augmentation method is proposed, called sliced mosaic permutation, which makes the original image features richer and alleviates the problem of category imbalance, via cutting the original images with different scales and recombining randomly. We apply SAEM and sliced mosaic permutation to the task of intracranial blood vessel segmentation, the result shows that our method outperforms competitive methods in both visualization results and quantitative evaluation.