ASS-GAN: Asymmetric semi-supervised GAN for breast ultrasound image segmentation

Ultrasound imaging is considered to be one of the important methods for diagnosing breast cancers, and lesion segmentation is an essential step in automatic computer-aided ultrasonic diagnosis. However, the high cost of ultrasound image labeling and the small amount of data in a single dataset hinder the progress of breast ultrasound (BUS) image segmentation algorithms. In this paper, we propose a novel asymmetric semi-supervised GAN (ASSGAN), which employs two generators and a discriminator for adversarial learning. The two generators can supervise each other, i.e., they can generate reliable segmentation predicted masks as guidance for each other without labels. Therefore, the unlabeled cases can be used to effectively promote model training. To verify the proposed method, we compared it with fully supervised and semi-supervised methods on three public BUS datasets (DBUI, OASBUI, SPDBUI) and one dataset (SDBUI) that we collected. DBUI, OASBUI, SPDBUI and SDBUI contain 647, 200, 320 and 1805 cases respectively. The experimental results show that the proposed method has excellent performance under the condition of having a small number of labeled images. Compared with fully supervised methods, our method is higher by 4.16% ti 13.94% in IoU. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

The study proposes a novel asymmetric semi-supervised GAN method for breast ultrasound image segmentation. The two generators can supervise each other and improve model training, resulting in excellent segmentation performance with a small number of labeled images.