BU-APSPNet: Breast Ultrasound Segmentation with Attention PSP-Unet Architecture

Segmentation of breast masses in breast ultrasound (BUS) examinations is important for the diagnosis of breast cancer (BC). For segmentation of breast tumor, we developed a deep learning (DL) model called BU-APSPNet. The model is trained and validated using the publicly available Breast Ultrasound Images Dataset (BUSI). The model's mean accuracy, dice score, IoU score, sensitivity, and specificity are 94.01%, 65.46%, 74.35%, 65.05%, 98.51 respectively. The performance of the model demonstrates a major improvement compared to the baseline model, with accuracy increasing by 4.35%, dice score by 52.4%, IoU score by 24.78%, and sensitivity by 55.47% respectively. Our new model effectively segments BUS masses with the ability to take into account multiple spatial scales, local context, and global context, with the provision of accurate pixel-level segmentation maps. The visual result demonstrates promising implications for more extensive dataset utilization.