CMAUNeXt: An Efficient Neural Network Based on Large Kernel and Multi-Dimensional Attention Module for Breast Tumor Segmentation

Advancements in deep learning technology have significantly contributed to the application of biomedical image analysis. Breast cancer, a disease commonly fatal in women, underscores the importance of early detection as a critical factor in improving survival rates. While ultrasound imaging (USG) has become the standard in preclinical screening, manual examination of USG images requires considerable time and cost, even for experienced radiologists. UNet and its variants have dominated the field of medical image segmentation with good performance. However, these models still need challenges, such as overly simplistic convolutional blocks with limitations in accessing global-scale information and skip connections that merely concatenate encoder and decoder features, proving less effective in suppressing irrelevant features. To address these challenges, this paper proposes a breast tumor segmentation network, CMAUNeXt, which combines short residual ConvNeXt (srCX) with a Multi-Dimensional Attention Module (MDAM). Specifically, we design the srCX block with depthwise convolution, large kernels, and two pointwise convolutions with an inverted bottleneck design to extract global context information efficiently. Meanwhile, the MDAM module is designed to strengthen valuable features and minimize less relevant information. According to experimental results on two breast ultrasound datasets, Dataset B and BUSI, the segmentation performance of the proposed CMAUNeXt model is superior, lighter, and has lower computational costs than previous models.