Cross-Tissue/Organ Transfer Learning for the Segmentation of Ultrasound Images Using Deep Residual U-Net

• Published in: Journal of medical and biological engineering Vol. 41; no. 2; pp. 137 - 145
• Main Authors: Huang, Haibo, Chen, Haobo, Xu, Haohao, Chen, Ying, Yu, Qihui, Cai, Yehua, Zhang, Qi
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01-04-2021; Springer Nature B.V
• Subjects: Biomedical Engineering and Bioengineering; Breast cancer; Cell Biology; Datasets; Engineering; Image processing; Image segmentation; Imaging; Knowledge management; Learning; Medical imaging; Original Article; Radiology; Transfer learning; Ultrasonic imaging; Ultrasonic testing; Ultrasound; Residual block; Image segmentation; Transfer learning; Ultrasound; U-net; Convolutional neural network
• ISSN: 1609-0985; 2199-4757
• DOI: 10.1007/s40846-020-00585-w

Purpose
Ultrasound image segmentation is a crucial step in computer-aided diagnosis. In this study, we propose a cross-tissue/organ segmentation method based on the transfer learning method and a modified deep residual U-Net model.
Methods
We present a modified deep residual U-Net model by integrating a U-Net architecture with residual blocks to leverage the advantages of both components. Next, we explore a cross-tissue/organ transfer learning method for ultrasound image segmentation, which transfers the knowledge of ultrasound image segmentation from one tissue/organ to another, e.g. from tendon images to breast tumor images and vice versa. We evaluated the proposed method by performing four groups of experiments on three medical ultrasound datasets, consisting of one tendon dataset and two breast datasets, along with one non-medical dataset.
Results
The results showed an overall performance improvement by our method in terms of the Dice coefficient and Jaccard index. It was demonstrated that our modified deep residual U-Net exceeded the standard U-Net and residual U-Net, and the cross-tissue/organ transfer learning was superior to training from scratch and to transfer learning between divergent domains.
Conclusion
Our method shows potential to accurately segment medical ultrasound images.