Breast Mass Detection in Mammograms via Blending Adversarial Learning

• Published in: Simulation and Synthesis in Medical Imaging Vol. 11827; pp. 52 - 61
• Main Authors: Lin, Chunze, Tang, Ruixiang, Lin, Darryl D., Liu, Langechuan, Lu, Jiwen, Chen, Yunqiang, Gao, Dashan, Zhou, Jie
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Adversarial deep learning; Digital mammography; Mammogram synthesis; Mass detection
• ISBN: 3030327779; 9783030327774
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-32778-1_6

Deep learning approaches have recently been proposed for breast cancer screening in mammograms. However, the performance of such deep models is often severely constrained by the limited size of publicly available mammography datasets and the imbalance of healthy and abnormal images. In this paper, we propose a blending adversarial learning method to address this issue by regularizing the imbalanced data with synthetically generated abnormal samples. Unlike most existing data generation methods that require large-scale training data, our approach is carefully designed for augmenting small datasets. Specifically, we train a generative model to simulate the growth of mass on normal tissue by blending mass patches into healthy breast images. The resulting synthetic images are exploited as complementary abnormal data to make the training of deep learning based mass detector more stable and the resulting model more robust. Experimental results on the commonly used INbreast dataset demonstrate the effectiveness of the proposed method.