# TransAttUnet: Multi-level Attention-guided U-Net with Transformer for Medical Image Segmentation

**Authors:** Bingzhi Chen, Yishu Liu, Zheng Zhang, Guangming Lu, David Zhang  
**arXiv:** [2107.05274](https://arxiv.org/abs/2107.05274)  
**Date:** 2021-07-12  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2107.05274v1/

## Key points

- Automatic medical image segmentation has advanced with deep encoder-decoder architectures and large-scale annotated datasets.
- Existing standard models face challenges in capturing global contextual feature dependencies.
- The paper introduces a framework called TransAttUnet for medical image semantic segmentation.
- TransAttUnet combines multi-level guided attention and multi-scale skip connections to enhance traditional U-shaped architecture.
- The framework incorporates a self-aware attention module that includes Transformer Self Attention (TSA) and Global Spatial Attention (GSA).
- Multi-scale skip connections are established between decoder blocks to aggregate different semantic-scale upsampling features.
- TransAttUnet effectively addresses the information recession problem, improving diagnostic sensitivity and segmentation quality in medical image analysis.
- Extensive experiments demonstrate that TransAttUnet consistently outperforms state-of-the-art baselines.
- The proposed framework holds promise for advancing automatic medical image segmentation in various imaging modalities.

## Summary

The field of automatic medical image segmentation has made significant advancements with the development of deep encoder-decoder architectures and large-scale annotated medical datasets. However, existing standard models face challenges in fully capturing global contextual feature dependencies due to the information recession problem caused by convolution layers and consecutive sampling operations. To address these challenges, this paper introduces a novel framework called TransAttUnet for medical image semantic segmentation. combines multi-level guided attention and multi-scale skip connections to enhance the functionality and flexibility of traditional U-shaped architecture. Inspired by Transformer, the framework incorporates a self-aware attention (SAA) module that includes both Transformer Self Attention (TSA) and Global Spatial Attention (GSA). This allows for effective learning of non-local interactions between encoder features. Additionally,  establishes multi-scale skip connections between decoder blocks to aggregate different semantic-scale upsampling features. This strengthens the representation ability of multi-scale context information, resulting in more discriminative features. By addressing the information recession problem,  effectively alleviates the loss of fine details and improves diagnostic sensitivity and segmentation quality in medical image analysis. Extensive experiments on multiple medical image segmentation datasets demonstrate that  consistently outperforms state-of-the-art baselines. The proposed framework holds promise for advancing automatic medical image segmentation and has potential applications in various imaging modalities.