DA-Slim UNETR: A Dual Attention Enhanced Network for Medical Image Segmentation

• Published in: 2024 IEEE 25th China Conference on System Simulation Technology and its Application (CCSSTA) pp. 310 - 314
• Main Authors: Jin, Jiawei, Yang, Sen, Tong, Jigang, Liu, Na
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21.07.2024
• Subjects: Accuracy; Biomedical imaging; Data models; Dual Attention; Feature extraction; Image segmentation; Medical image segmentation; Surveillance; Systems simulation; Three-dimensional displays; Transformer; Transformers; Tumors; U-net
• DOI: 10.1109/CCSSTA62096.2024.10691759

Image segmentation, a technique for classifying regions of interest at the pixel level, is essential in computer vision applications such as autonomous driving and surveillance. Slim UNETR is an efficient deep learning method for medical image segmentation, capable of processing 3D medical data to identify regions of interest, like tumors, for doctors. This network employs a U-shaped architecture and is categorized as a hybrid Transformer network, achieving high Dice accuracy for segmentation tasks. However, its skip connections inadequately extract positional and channel information from feature maps. To address this issue, we introduced DA-Black, a dual attention module that combines positional and channel attention. By integrating DA-Black, we developed DA-Slim UNETR. Slim UNETR achieved an accuracy of 80.27% on the MSD2019 Task-01. With DA-Black incorporated into the skip connections and bottleneck layers, DA-Slim UNETR's Dice segmentation accuracy improved to 81.54%, a 1.27% increase. This performance is also 2.38% higher than the mainstream medical segmentation network Swin UNETR.