SAASU‐UNet++: A Hybrid Swin Transformer–Based Deep Learning Model for Robust Breast Tumour Segmentation in Ultrasound Images

ABSTRACT Breast cancer is one of the most prevalent and life‐threatening diseases affecting women worldwide, where early detection is crucial for improving survival rates. Ultrasound imaging is widely used due to its non‐invasive, cost‐effective nature and suitability for dense breast tissues. This study introduces Swin‐attention‐enhanced atrous spatial pyramid pooling (ASPP)‐attention‐squeeze‐and‐excitation (SE) uncertainty‐aware U‐Net++ (SAASU‐UNet++), a novel hybrid segmentation framework designed for precise breast tumour segmentation in ultrasound images. The model incorporates a Swin Transformer encoder to capture global context, an attention‐enhanced ASPP module for multi‐scale feature extraction, and a hybrid U‐Net++ decoder embedded with SE blocks for channel recalibration. To improve robustness and uncertainty awareness, Monte Carlo dropout is applied. Training was performed with the Adam optimiser (batch size 32, 100 epochs) on the BUSI dataset containing 780 annotated images (normal, benign, and malignant). Images were resized to 224 × 224, normalised, and augmented with flipping, rotation, and zooming. The proposed model achieved a Dice Coefficient of 96.25% and a Jaccard Index of 92.34%, outperforming state‐of‐the‐art approaches. Ablation studies highlighted the effectiveness of each component, and visual results confirmed close alignment with ground truth, even in challenging cases. SAASU‐UNet++ demonstrates strong potential for real‐world clinical integration in automated breast cancer diagnosis.