A Novel Multi-Scale Boundary Guide Model for Ultrasound Breast Lesion Segmentation

Abstract

Breast carcinoma affects over 2.26 million individuals annually, emphasizing the urgent need for early and accurate detection methods to improve clinical outcomes. While ultrasound imaging is widely used for its non-radiative nature, challenges such as poor contrast and complex lesion structures reduce its diagnostic efficacy. This study presents a robust automated system for breast lesion segmentation in ultrasound images using advanced deep learning techniques to enhance detection accuracy and reliability. The proposed framework integrates an encoder-decoder architecture with spatial attention mechanisms and boundary segmentation strategies. A Shearlet-based anisotropic diffusion technique is employed for preprocessing to enhance image quality. For feature extraction, the system utilizes Inception v4 and a residual multi-scale, multilevel spatial attention model. Segmentation is performed using a multi-scale boundary-guided, dual-resolution, lightweight triangulation topology aggregation approach. Additionally, Explainable AI (XAI) techniques such as SHAP and LIME are leveraged to identify genes crucial to various cancer stages. SHAP highlights top-ranking genes across the genomic dataset, while LIME offers instance-level insights. Influential genes such as PLA2G10, MST1R, F13B, and CAMK1, identified by the proposed GeneXAI model, align with biomarkers recognized in cutting-edge biological research. The model also classifies these genes as prognostic or non-prognostic based on clinical significance. The proposed system achieved a Jaccard index of 87.09%, Dice coefficient of 98.63%, precision of 95.51%, and recall of 92.78% across benchmark datasets. These results affirm the model’s potential in supporting early breast cancer detection and aiding clinicians in timely, informed decision-making.