Fetal-BCP: Addressing Empirical Distribution Gap in Semi-Supervised Fetal Ultrasound Segmentation
Pham, Ha-Hieu Khanh, Le Tran Quoc Nguyen, Hoang-Thien Vu, Nguyen Lan Vi Dinh, Quang-Vinh Nguyen, Thanh-Huy Li, Xingjian Xu, Min
Pham, Ha-Hieu
Khanh, Le Tran Quoc
Nguyen, Hoang-Thien
Vu, Nguyen Lan Vi
Dinh, Quang-Vinh
Nguyen, Thanh-Huy
Li, Xingjian
Xu, Min
Supervisor
Department
Computer Vision
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Fetal ultrasound imaging is critical for examining cervical architecture, but segmentation remains difficult due to hard-to-learn features under a lack of labeled data scenarios. In this paper, we present a unique semi-supervised system for cervical segmentation that efficiently uses both labeled and unlabeled images. Inspired by Bidirectional Copy-Paste (BCP), our proposed method, named Fetal-BCP, uses a Mean Teacher framework and thorough data augmentation approaches to reduce the distribution mismatch between labeled and unlabeled data. Our approach significantly lessens the human annotation effort by balancing trustworthy supervision with inferred annotations through consistency regularization and pseudo-labeling. Our results maintained the top rank on the leaderboard of the Fetal Ultrasound Grand Challenge of ISBI 2025 Challenge, surpassing many state-of-the-art methods on both Dice and HD metrics.
Citation
H. -H. Pham et al., "Fetal-BCP: Addressing Empirical Distribution Gap in Semi-Supervised Fetal Ultrasound Segmentation," 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), Houston, TX, USA, 2025, pp. 1-4, doi: 10.1109/ISBI60581.2025.10980925.
Source
2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)
Conference
IEEE International Symposium on Biomedical Imaging
Keywords
Ha-Hieu Pham, Le Tran Quoc Khanh, Hoang-Thien Nguyen, Nguyen Lan Vi Vu, Quang-Vinh Dinh, Thanh-Huy Nguyen
Subjects
Source
IEEE International Symposium on Biomedical Imaging
Publisher
IEEE