ATFF: Advanced Transformer with Multiscale Contextual Fusion for Medical Image Segmentation

Abstract

Deep convolutional neural network (CNN) greatly promotes the automatic segmentation of medical images. However, due to the inherent properties of convolution operations, CNN usually cannot establish long-distance interdependence, which limits the segmentation performance. Transformer has been successfully applied to various computer vision, using self-attention mechanism to simulate long-distance interaction, so as to capture global information. However, self-attention lacks spatial location and high-performance computing. In order to solve the above problems, we develop a new medical transformer, which has a multi-scale context fusion function and can be used for medical image segmentation. The proposed model combines convolution operation and attention mechanism to form a u-shaped framework, which can capture both local and global information. First, the traditional converter module is improved to an advanced converter module, which uses post-layer normalization to obtain mild activation values, and uses scaled cosine attention with a moving window to obtain accurate spatial information. Secondly, we also introduce a deep supervision strategy to guide the model to fuse multi-scale feature information. It further enables the proposed model to effectively propagate feature information across layers, Thanks to this, it can achieve better segmentation performance while being more robust and efficient. The proposed model is evaluated on multiple medical image segmentation datasets. Experimental results demonstrate that the proposed model achieves better performance on a challenging dataset (ETIS) compared to existing methods that rely only on convolutional neural networks, transformers, or a combination of both. The mDice and mIou indicators increased by 2.74% and 3.3% respectively.