Mechanisms and Applications of Attention in Medical Image Segmentation: A Review

Abstract

The core task of medical image segmentation based on deep learning is to quickly obtain good results through low-cost auxiliary modules. The attention mechanism, relying on the interacting features of the neural network, is one of the lightweight schemes to focus on key features, which is inspired by the characteristics of selective filtering information in human vision. Through the investigation and analysis, this paper argues that the common attentional mechanisms can be mainly classified into four types according to their structure and form: (i) conventional attention based on feature interaction, (ii) multi-scale/multi-branch-based attention, (iii) Self-similarity attention based on key-value pair queries, (iv) hard attention, etc.
 Medical images contain poor and blur descriptions of contextual information than natural images. They are usually re-imaging by the feedback intensity of the medium signal since most of them have low contrast and uneven appearance, as well as contain noise and artifacts. In models based on deep learning, without the ability to focus on key descriptive information or features, it is difficult for well-designed models to perform theoretically. This paper shows that attention mechanisms can guide downstream medical image analysis tasks to master discernible expected features while filtering and suppressing irrelevant information to enhance the intensity of target features. Therefore, the network performance can be improved through continuous highly accurate feature spatial evolution.