Weakly Supervised Erase to Segment Nodules in Breast Ultrasound

View the PDF file from the paper entitled Flip Learning: Ending a weak supervision to cut nodules in the ultrasound of the breast, by Yuhao Huang and 10 other authors

PDF HTML (experimental) view

a summary:The exact division of nodules in the two -dimensional ultrasound (BUS) and 3D automatic ultrasound (ABUS) is very important for clinical diagnosis and treatment planning. Therefore, the development of an automatic system of fragmentation of nodules can enhance the independence of the user and accelerate the clinical analysis. Unlike the fully supervised learning, retail with weak supervision (WSS) can simplify the process of arduous and complex illustrations. However, the current WSS methods face challenges in achieving accurate nodules, as many of them depend on inaccurate activation maps or immune mask algorithms. In this study, we offer a new -agent -based WSS Work frame called Flip Learning, which depends only on two -dimensional/3D boxes for accurate retail. Specifically, multiple factors are used to erase the target of the square to facilitate the stirring of the classification mark, with the area of ​​the area as an expected segmentation mask. The main contributions of this research are as follows: (1) Approval of a supervoxel approach to encrypting the unified environment, capturing the border borders and accelerating the learning process. (2) Entering three accurate rewards, which include a bonus of classification and severity distribution bonuses, to directly direct the agents erasing, thus avoiding both division and excessive assembly. (3) Implementing a curriculum learning strategy to enable agents to gradually interact with the environment, thus enhancing learning efficiency. It is widely validated on large bus groups in buses and ABUS, our learning method outperforms modern WSS styles and basic models, and achieves comparable performance as educational algorithms subject to fully overseeing.