A Neural Ordinary Differential Equations (ODE)-Based Model with Attention Mechanism for Tumor Segmentation in Breast Ultrasound Images

Breast cancer is fairly common among the women worldwide, and ultrasound for screening and assessing breast masses is one of the most frequently used approaches due to its advantages like portable, noninvasive and easy to operate. The automatic tumor segmentation in breast ultrasound images is conducive to clinical diagnosis and treatment. Deep neural network for segmentation has become a popular tool nowadays to assist doctors and provide the references but it still faces some challenges such as large amounts of parameters, lack of interpretability, overfitting problem, etc. In this paper, we propose a segmentation model named Att-U-Node which combines a neural ODE-based framework U-Node and an attention-based module CBAM, trying to alleviate the problems mentioned above. Three public available breast ultrasound image datasets including BUSI, BUS and OASBUD as well as a private breast DCE-MRI dataset are used to measure the performance of the proposed model and the related neural networks and then we further extend experiments to 3D scenes for tumor segmentation with the data selected from QIN Breast DCE-MRI. The experimental results show that our proposed Att-U-Node outperforms the comparison methods with respect to the most of the metrics while mitigates the common problems of deep neural network.