DenseATT-Net: Densely-Connected Neural Network with Intensive Attention Modules for 3D ABUS Mass Segmentation
Hengyu Zhang, Jingxuan Xu, Mengyu Wang, Yanfeng Li · 2022 7th International Conference on Image, Vision and Computing (ICIVC) · 2022
Accurate segmentation of breast mass in 3D automated breast ultrasound (ABUS) images is important in breast cancer analysis. However, it is hard to obtain enough labeled ABUS images for training segmentation networks, which may lead to over-fitting problem in deep learning based methods. Aiming at this problem, a lightweight segmentation network D2U-Net is selected as the baseline. ABUS images have a low signal-to-noise ratio and serious artifacts, which makes mass boundary unclear. To address this problem, different kinds of attention modules are inserted into the segmentation network. These attention modules include spatial attention, channel attention, convolutional block attention module (CBAM) and squeeze-and-excitation (SE) block. The whole segmentation network is termed as DenseATT-Net. An ABUS dataset with 170 volumes is employed to verify the segmentation performance. Experimental results show that the proposed method performs better than other segmentation models on 3D ABUS images.