SOTS: Leveraging Self-supervised Pretraining for Label-Efficient Ovarian Tumor Segmentation in Ultrasound Images

Ovarian cancer is one of the most serious diseases globally, and ultrasound imaging is a widely used modality for its diagnosis and monitoring. Accurate segmentation of ovarian tumors is critical for reliable morphological assessment and clinical decision-making, yet supervised models are often limited in performance when annotated data are scarce due to the high cost and expertise required for labeling. To address this challenge, we propose a Self-supervised Ovarian Tumor Segmentation (SOTS) framework that leverages unlabeled ultrasound images to learn robust feature representations. In the proposed framework, for the supervised stage, a new architecture named SovaSegNet-U is proposed. The encoder is pretrained using the Barlow Twins objective to learn invariant and non-redundant embeddings, while an uncertainty-based augmentation strategy further strengthens representation robustness by prioritizing challenging transformations during pretraining on unlabeled data. The pretrained encoder is subsequently fine-tuned under varying labeled data ratios, consistently outperforming purely supervised models in low-label scenarios. Notably, SOTS achieves segmentation accuracy comparable to fully supervised U-Net variants while requiring substantially fewer annotations. These results demonstrate that the proposed approach provides an annotation-efficient solution for ovarian tumor segmentation and holds strong potential for supporting clinical decision-making. Source code is available at https://github.com/SonBH0410/SOTS .