CSwin-PNet: A CNN-Swin Transformer combined pyramid network for breast lesion segmentation in ultrasound images
H Yang, D Yang - Expert Systems with Applications, 2023 - Elsevier
H Yang, D Yang
Expert Systems with Applications, 2023•ElsevierCurrently, the automatic segmentation of breast tumors based on breast ultrasound (BUS)
images is still a challenging task. Most lesion segmentation methods are implemented
based on a convolutional neural network (CNN), which has limitations in establishing long-
range dependencies and obtaining global context information. Recently, transformer-based
models have been widely used in computer vision tasks to build long-range contextual
information due to their powerful self-attention mechanism, and their effect is better than that …
images is still a challenging task. Most lesion segmentation methods are implemented
based on a convolutional neural network (CNN), which has limitations in establishing long-
range dependencies and obtaining global context information. Recently, transformer-based
models have been widely used in computer vision tasks to build long-range contextual
information due to their powerful self-attention mechanism, and their effect is better than that …
Abstract
Currently, the automatic segmentation of breast tumors based on breast ultrasound (BUS) images is still a challenging task. Most lesion segmentation methods are implemented based on a convolutional neural network (CNN), which has limitations in establishing long-range dependencies and obtaining global context information. Recently, transformer-based models have been widely used in computer vision tasks to build long-range contextual information due to their powerful self-attention mechanism, and their effect is better than that of a traditional CNN. In this paper, a CNN and a Swin Transformer are linked as a feature extraction backbone to build a pyramid structure network for feature encoding and decoding. First, we design an interactive channel attention (ICA) module using channel-wise attention to emphasize important feature regions. Second, we develop a supplementary feature fusion (SFF) module based on the gating mechanism. The SFF module can supplement the features during feature fusion and improve the performance of breast lesion segmentation. Finally, we adopt a boundary detection (BD) module to pay additional attention to the boundary information of breast lesions to improve the boundary quality in the segmentation results. Experimental results show that our network outperforms state-of-the-art image segmentation methods on breast ultrasound lesion segmentation.
Elsevier