Breast Ultrasound Image Segmentation with Multidimensional Attention and Spatial-Channel Squeeze-and-Excitation Block

Breast cancer remains one of the most common and deadly cancers worldwide, with over 2.2 million new cases and more than 666,000 deaths recorded in 2022. Early diagnosis is essential for increasing survival rates, as timely treatment can successfully cure <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$70-80 \%$</tex> of early-stage cases. Ultrasound imaging is a valuable diagnostic tool, particularly Automated Breast Ultrasound (ABUS). However, it is often time-consuming and reliant on operator expertise. This research proposes an improved breast cancer detection approach by developing a deep learning model based on Half-UNet, incorporating multidimensional attention (MA) and a Spatial-Channel Squeeze-and-Excitation (scSE) Block. The enhanced model integrates multidimensional attention, which consists of three attention mechanisms: channel attention (CA), spatial attention (SA), and pixel attention (PA). Evaluated on the Breast Ultrasound Image (BUSI) dataset, the model achieves a Dice Score of 81.40 %, an IoU of 71.52%, a Specificity of 87.02%, and a Sensitivity of 80.28 %, with a total of 0.19 million parameters. Designed for accurate and computationally efficient breast cancer segmentation, this model aims to support early diagnosis and improve clinical outcomes.