Go to MICCAI 2025 Challenge IUGC homepageNoisy Student-Based Self-Training Enhances Landmark Detection in Intrapartum Ultrasound
Keywords: Anatomical landmark detection, Intrapartum ultrasound, Semi-supervised learning, Noisy Student framework
TL;DR: This paper proposes a Noisy Student-based semi-supervised method using limited labeled and partial unlabeled images to improve intrapartum ultrasound landmark detection accuracy and robustness.
Abstract: Accurate detection of fetal anatomical landmarks in ultrasound images during labor is crucial for clinical labor assessment. Despite significant progress in deep learning for medical image analysis, achieving high-precision and robust keypoint detection remains challenging under the realistic condition of scarce annotated data. Inspired by the Noisy Student paradigm in image classification, this paper proposes an improved semi-supervised method tailored for keypoint detection tasks.We construct a DenseUNet teacher-student framework to perform collaborative training using limited annotated data and a portion of unlabeled images. Specifically, the teacher model is trained on the labeled set to generate heatmap pseudo-labels for the unlabeled data; the student model, supervised by the pseudo-labels, leverages the dense connectivity of DenseNet to enhance feature reuse and gradient flow, and incorporates Dropout in the decoder to improve robustness. Furthermore, a linearly-decayed MixUp strategy is adopted for input perturbation, combined with heatmap supervision, to achieve a smooth transition from strong perturbation training to stable convergence. Experiments on the IUGC 2025 test set demonstrate that the proposed method significantly improves landmark detection performance, achieving an average distance error (Distance) of 13.1574 and an AOP\_MAE score of 4.4244, which verifies the effectiveness of the method in scenarios with limited annotation resources.
The project source code is available at: https://github.com/apuomline/IUGC2025.
Submission Number: 3
Loading