由于乳腺癌的高发生率和高死亡率以及亚洲女性乳腺周围腺体遮蔽和病变组织重叠的影响在乳腺x光片中对肿块进行分割是很重要的很困难的。在乳腺x光片肿块分割中将患者同侧俩不同角度的视图cc view和MLO view信息整合起来进行建模是特别重要。然而大多数现有的方法都是将患者同侧俩不同的视图视为独立的个体送入网络中或者只是将不同的视图特征进行简单的连接这些方法都比较粗糙。在本文中我们提出了一个基于UNet

Due to the high incidence and mortality rates of breast cancer, as well as the effect of breast tissue overlapping and obscuring in Asian women, segmenting masses in mammograms is crucial but challenging. In mammographic mass segmentation, it is particularly important to model the two different views of the same breast (CC view and MLO view) jointly. However, most existing methods treat the two views as independent entities or simply concatenate the features from different views, which are rather crude. In this paper, we propose a UNet and transformer-based architecture called xxxnet to learn the corresponding masses in the two views for segmentation. In xxxnet, we introduce the xx module to achieve dynamic interaction between masses in CC and MLO views, where the proposed complementary attention mechanism plays a crucial role. We also introduce relative positional encoding specifically designed for the two views of the breast, in order to locate the masses in CC and MLO views in the early stage of training. The combination of these two designs achieves accurate segmentation of mammographic masses. Experimental results show that xxxnet outperforms state-of-the-art methods on our in-house dataset, and its performance improvement is more significant when the amount of training data is slightly reduced