COMT: Co-Training Mean Teachers Semi-Supervised Training Framework for Cervical Segmentation
Yu Chen, Bo Deng, Zilun Peng, Yongle Zhou, Pan Sun, Jin Wan, Jieyun Bai
Abstract
Cervical morphology is essential for evaluating gynecological health, and accurate segmentation of cervical muscles in transvaginal ultrasound (TVUS) images is crucial for analyzing cervical anatomy and function. However, the low resolution and noise in ultrasound images make manual labeling time-consuming. Existing semi-supervised medical image segmentation methods primarily focus on generating additional supervisory signals from unlabeled images. This paper proposes a semi-supervised deep learning framework based on MeanTeacher, which utilizes different model architectures to enhance diversity in unlabeled images and applies minimum entropy to reconstruct pseudo-soft labels. Experiments on the ISBI FUGC 2025 dataset demonstrate the effectiveness of this method, contributing to early risk prediction and personalized treatment in clinical decision-making.