MSDAHNet: A multi-scale dual attention hybrid convolution network for breast tumor segmentation

Abstract

Accurate breast tumor segmentation is crucial for physicians in breast cancer diagnosis and clinical analysis. To address challenges such as the diversity in tumor size and shape, as well as the indistinct boundaries in breast tumor segmentation, we propose a multi-scale dual attention hybrid convolution network (MSDAHNet). MSDAHNet incorporates a residual dual channel attention module in the encoder and decoder, allowing it to capture essential image features more effectively. In the down-sampling of encoding, both max-pooling and average-pooling are utilized to minimize the loss of fine features. Besides, the network introduces a multi-scale feature pyramid aggregation module and a multi-scale hybrid convolution module to facilitate the integration of important feature mappings and enhance the tumor regions. We conduct experiments using four breast imaging datasets. Specifically, Dice scores of MSDAHNet on BUSBRA, BUET_BUSD, private hospital, and BreaDM datasets are 0.906, 0.842, 0.853, and 0.832, respectively, and Iou scores are 0.841, 0.768, 0.780, and 0.745 for each dataset respectively. The experimental results demonstrate that MSDAHNet outperforms comparing to existing state-of-the-art methods, showing good potential in breast tumor segmentation.