A semi-supervised approach for breast tumor segmentation using sparse transformer attention UNet

Abstract

Accurate segmentation of breast tumors, especially in younger women, remains a significant challenge in cancer research. Ultrasound imaging, a non-invasive screening method, relies on tumor characteristics such as size and texture, which are crucial for clinicians to make precise diagnoses. However, the lack of annotated datasets necessitates the development of advanced deep learning models. While traditional U-Net models, based on Convolutional Neural Networks (CNNs), excel at local feature extraction, they struggle to capture long-range dependencies. Transformer models address this limitation but are computationally demanding and require large, annotated dataset. To overcome these challenges, we propose a semi-supervised learning approach with three components: The Diverse Image Generation Network (DIGN), the Adaptive Probability Mapping Network (APMN), and STA-UNet (Sparse Transformer Attention UNet), a novel architecture designed to efficiently capture long-range dependencies while reducing computational cost. Experimental results demonstrate that STA-UNet significantly outperforms traditional U-Net models. On the Mendeley dataset, STA-UNet achieves a 4.10 % improvement in the Jaccard Similarity Coefficient (JSC), a 3.84 % increase in the Dice Similarity Coefficient (DSC), and a 38.00 % reduction in Hausdorff Distance (HD). Similarly, on the SIIT dataset, STA-UNet shows a 1.92 % increase in JSC, a 2.02 % improvement in DSC, and a 30.58 % reduction in HD.