Joint-phase attention network for breast cancer segmentation in DCE-MRI

Breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) plays an important role in the screening and treatment evaluation of high-risk breast cancer. The segmentation of cancerous regions is an essential step for the comprehensive analysis of breast MRI. Nevertheless, automatic and robust segmentation is still very challenging because of the large differences of cancers in size, morphology, and intensity appearance. To alleviate these issues, we propose a simple yet effective two-stage approach, which simultaneously exploits pre- and post-contrast enhanced information to segment the breast cancer. In particular, we first offer a breast segmentation network to predict the breast region of interest (ROI) and therefore excluding confounding information from thorax region in the whole MRI scans. Moreover, inspired by the radiologists' examination routine which takes full advantage of the MRI sequences to make the diagnosis, we suggest a joint-phase attention network in order to mine both pre- and post-contrast representations for the segmentation of cancerous regions. The accuracy and generalizability of the proposed network is validated on our collected DCE-MRI dataset containing 550 subjects (with 748 biopsy-proven breast cancers) from 3 different centers (one as internal dataset and two as external datasets). The primary evaluation metrics are Dice similarity coefficient (Dice), Jaccard index (Jaccard), and average symmetric surface distance (ASSD). Our network consistently achieves satisfactory segmentation results, by generating an average Dice of 88.77%/82.77%/83.03%, Jaccard of 81.27%/71.89%/73.23%, and ASSD of 2.21/3.63/2.69, on one internal and two external datasets, respectively. Our method offers an effective cancer segmentation approach for the breast DCE-MRI examination. The code is publicly available at https://github.com/ryandok/JPA.