UJAT-Net: A U-Net Combined Joint-Attention and Transformer for Breast Tubule Segmentation in H&E Stained Images
Li-Wang Liu,
Zhao Huang,
Kao-Yan Lu,
Zu-Xuan Wang,
Yao-Ming Liang,
Shi-Yu Lin,
Yan-Hong Ji
Affiliations
Li-Wang Liu
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
Kao-Yan Lu
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
Zu-Xuan Wang
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
Yao-Ming Liang
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
Key Laboratory of Atomic and Subatomic Structure and Quantum Control (Ministry of Education), Guangdong Basic Research Center of Excellence for Structure and Fundamental Interactions of Matter, School of Physics, South China Normal University, Guangzhou, China
The formation of breast tubules is an important evaluation index in the pathological grading of breast cancer. However, the tubules of breast present a wide variety of morphologies and a significant demonstrated significant advantages in the automatic analysis of histopathology images. We propose a Joint Attention and Transformer U-Net network to accurately segment breast tubules, named UJAT-Net. UJAT-Net uses the Joint Attention Block (named JA BLOCK) as the encoder of the network to enhance the extraction effect of the network for different layer features. And the Channel Cross fusion with Transformer (named CCT) module is used as the skip connection structure of the network. Furthermore, we employ a Transpose Cross Attention (named TCA) module as the decoder of the network to fuse the features of the skip connection layer and the decoder. Experimental results on our own invasive breast cancer tubule (Tubule of Breast Cancer, TBC) dataset and the benchmark dataset (Glas) of the GlaS challenge contest at MICCAI’2015 demonstrate that our method achieves competitive performance.
