Difficulty-aware prior-guided hierarchical network for adaptive segmentation of breast tumors

Breast tumor segmentation is vital to tumor detection at the early stages. Deep learning methods are typically used in automatic tumor segmentation tasks. However, in existing methods, the difference between pixels is disregarded, and the union network architecture is used to segment all pixels; these methods involve a tradeoff between accuracy and efficiency. A novel, difficulty-aware, prior-guided hierarchical network for the adaptive segmentation of breast tumors is presented herein. A difficulty prior learning module is proposed to learn the pixel’s difficulty prior to guild adaptive segmentation in the proposed network. To achieve a more accurate segmentation of hard pixels, a hard pixel processing unit is presented to learn more discriminative features for hard pixels. Experiments are conducted based on three datasets. The experimental results show that the proposed methods outperform traditional deep learning methods and achieve a balance between accuracy and efficiency.