Weakly and semi-supervised segmentation of uterus using coarse annotations in MRI

Endometrial cancer (EC) seriously threatens women's lives via invasion and metastasis, potentially causing multi-system organ failure. Accurate segmentation of the uterus on MR images has important implications for determining the depth of myometrial infiltration in EC, and for benign uterine diseases such as uterine fibroids, it enables efficient quantification of uterine volume and clear visualization of anatomical contours, thereby providing essential support for clinical disease assessment and preoperative planning. However, current deep learning-based uterine segmentation relies on manual pixel-level labeling, which is time-consuming and subjective. This study proposes a new uterus segmentation framework based on a combination of weakly supervised and semi-supervised learning aimed at learning from a small amount of scribble-labeled data and a large amount of unlabeled data. Specifically, the framework includes a two-branch network with residual blocks and two different perturbations. On the other hand, there may be some incorrectly predicted pixels in the pseudo labels. In this paper, a confidence-guided strategy is used to filter out the incorrect pixels in the pseudo labels to improve the network segmentation performance. The proposed method is validated on MR images of 220 patients with uterine diseases. The results show that the proposed method outperforms existing weakly supervised and semi-supervised segmentation methods, especially when the labeled ratio is small. In addition, the proposed method achieves comparable performance to fully supervised methods for any labeling rate.