Ovarian Ultrasound Image Segmentation with Limited Training Data

Tóm tắt

Ultrasound imaging is pivotal for ovarian tumor diagnosis, yet it poses significant segmentation challenges due to severe speckle noise, low contrast, and high inter-patient morphological variability. The limited availability of annotated medical data, making few-shot segmentation a practical and necessary solution. Although recent universal few-shot models such as UniverSeg demonstrate promising cross-task generalization, their performance is highly sensitive to stochastic support set sampling, often leading to unstable and inferior predictions. To address this limitation, we propose a CLIP-guided support retrieval strategy that replaces random sampling with deterministic similarity-based selection in the semantic embedding space of the Contrastive Language–Image Pre-training model. By retrieving morphologically consistent support samples for each query image, our method enhances structural alignment and reduces support-query mismatch without requiring additional training or model modification. Extensive experiments on two ovarian ultrasound datasets, OvaTUS and OTU_2D, demonstrate that our approach consistently outperforms the baseline UniverSeg and other few-shot methods. Specifically, on the OvaTUS dataset, our method achieves a Dice Coefficient of 75.84% and Intersection over Union of 64.87%, surpassing the random selection baseline by 2.19% and 2.69%, respectively. Furthermore, our approach shows superior robustness in extreme few-shot settings (N = 1), improving the Dice score by over 8% compared to the baseline. Code will be publicly released upon acceptance.