SHA-MTL: soft and hard attention multi-task learning for automated breast cancer ultrasound image segmentation and classification

Purpose The automatic analysis of ultrasound images facilitates the diagnosis of breast cancer effectively and objectively. However, due to the characteristics of ultrasound images, it is still a challenging task to achieve analyzation automatically. We suppose that the algorithm will extract lesion regions and distinguish categories easily if it is guided to focus on the lesion regions.Method We propose a multi-task learning (SHA-MTL) model based on soft and hard attention mechanisms for breast ultrasound (BUS) image simultaneous segmentation and binary classification. The SHA-MTL model consists of a dense CNN encoder and an upsampling decoder, which are connected by attention-gated (AG) units with soft attention mechanism. Cross-validation experiments are performed on BUS datasets with category and mask labels, and multiple comprehensive analyses are performed on the two tasks.Results We assess the SHA-MTL model on a public BUS image dataset. For the segmentation task, the sensitivity and DICE of the SHA-MTL model to the lesion regions increased by 2.27% and 1.19% compared with the single task model, respectively. The classification accuracy and F1 score increased by 2.45% and 3.82%, respectively.Conclusion The results validate the effectiveness of our model and indicate that the SHA-MTL model requires less a priori knowledge to achieve better results by comparing with other recent models. Therefore, we can draw the conclusion that paying more attention to the lesion region of BUS is conducive to the discrimination of lesion types.