Home | Help Center

Endless possibilities in academia

Application of U-Net and its variants in ultrasound image segmentation

Yuxiang Wang1, Miao Zhou2, Fangfang Chen1, Jintao Duan1, Liangqing Lin3, Qinghua Wu3, Wenhui Guo4, Haipo Cui1


1School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China. 2Jiangsu Cancer Hospital, Nanjing 213164, China. 3Anesthesiology, The First Hospital of Putian, Putian 351100, China. 4School of Anesthesiology, Second Military Medical University/Naval Medical University, Shanghai 200433, China.


Address correspondence to: Haipo Cui, School of Health Science and Engineering, University of Shanghai for Science and Technology, No. 516 Jungong Road, Yangpu District, Shanghai 200093, China. E-mail: h_b_cui@163.com.


Received September 27, 2024; Accepted January 22, 2025; Published April 1, 2025


DOI: https://doi.org/10.61189/861515qdddmg


Highlights

● This review provides a comprehensive overview of various U-Net architectures and their variants, including the original U-Net, U-Net++, Attention U-Net, and ResU-Net, along with a discussion on potential improvements to the

U- Net architecture.

● The differences between these network architectures are analyzed in terms of training complexity and computa-   tional requirements.

● The review delves into the application of U-Net and its variants in ultrasound imaging, discussing both the advan-   tages and limitations of each model in various ultrasound contexts. Relevant literature on the application of each network architecture in ultrasound is also summarized.

Abstract

Ultrasonography plays an important role in the fields of obstetrics, gynecology, cardiology, and hepatology, as well as ultrasound-guided nerve blocks, interventional therapy, and surgical navigation due to its non-invasive, real-time imaging and radiation-free characteristics. Recently, with the advancement of artificial intelligence, ma-chine learning and deep learning algorithms have brought significant innovations to ultrasound imaging technology in the medical field. U-Net is widely recognized as one of the most commonly used deep learning models in medi-cal image processing. This paper explores the application of the U-Net family of models in ultrasound imaging. The network architecture of the original U-Net, comprising encoder and decoder components, is first delineated. Next, classical variants, such as U-Net++, Attention U-Net, and ResU-Net, are introduced. The application of U-Net mod-els in ultrasound and their segmentation performance are then reviewed, with Dice coefficients highlighted as the primary evaluation metric. Finally, the paper provides a comparative analysis of the advantages and disadvantag-es of the U-Net family of models.

Keywords: U-Net, U-Net++, Attention U-Net, ResU-Net, ultrasonography, segmentation performance

Wang YX, Zhou M, Chen FF, Duan JT, Lin LQ, Wu QH, Guo WH, Cui HP. Application of U-Net and its variants in ultrasound image segmentation. Med Artif Intell 2025 Apr; 1(1): 27-38.

Copy