Segmentation of Tumor Regions in BUSI Breast Ultrasound Images Based on DRA-UNet Model with CBAM

Abstract

Breast cancer remains a significant global health problem, highlighting the need for early screening and accurate diagnosis. Considering the complexity of ultrasound images, lack of clear boundaries, and variability in tumor shape and texture, overly simple models cannot meet the high precision requirements for lesion segmentation. In order to achieve the stability of automatic segmentation of breast lesions in ultrasound images, this paper proposes an attention-based U-Net framework (DRA-CBAM-UNet model), which integrates existing deep learning models. This includes using densely connected networks for feature reuse, adding residual connections to UNet, a residual feature extraction networks to reduce the difference between encoder and decoder feature maps, and combining dilated convolutions with convolutional block attention module (CBAM) attention mechanism to enhance the model's ability to capture local features. This paper uses the publicly available BUSI dataset to train and test the model, and experimental results show that our proposed method achieves higher segmentation accuracy compared to other reported methods.