Anatosegnet: Anatomy Based CNN-Transformer Network for Enhanced Breast Ultrasound Image Segmentation

Accurate segmentation of breast tumor boundaries is essential for effective breast cancer diagnosis. Many convolutional and transformer-based models have been proposed for the semantic segmentation of Breast UltraSound (BUS) images. However, transformer-based segmentation models are challenging to train on small medical datasets, and breast anatomical information is rarely incorporated into these models to enhance their performance. In this study, we propose AnatoSegNet, a novel hybrid network that integrates a CNN-based U-shaped architecture with a novel breast Anatomical Attention Module for BUS image segmentation. The proposed attention module introduces a novel differential transformer and a bias matrix that emphasizes the layer structure of BUS images while capturing long-range dependencies, thereby improving the network's feature extraction capabilities. The proposed model is evaluated on two public BUS image datasets and achieves superior tumor IoU and F1 scores compared to state-of-the-art methods. The code is available at https://github.com/kuanhuang0624/AnatoSegNet.