3D Deep Adaptively Efficient Attention Model to segment Automated Breast Ultrasound

3D automated breast ultrasound (ABUS) data includes a huge number of slices with lesion-related artifacts and size variants, causing false positive cases. In addition, applying transformer-based models with self-attention considering long-range dependency to 3D ABUS data is challenging because of the small number of data. We propose 3D deep adaptively efficient attention model with a multi-task learning to segment the small number of 3D ABUS datasets. Our network consists of adaptive hybrid-transformer blocks for computationally efficient self-attention and multi-task learning to minimize false positive cases. Additionally, a novel attention gate is also implanted to considering long-range dependency between channel-attentive features. Our model performed 58.27%, 23.59 mm, and 0.054% on Dice, HD95, and FPR, respectively.