Text-Guided Cross Pseudo Supervision for Semi-Supervised Multi-Organ Segmentation in Abdominal CT Scans

Semi-supervised multi-organ segmentation in abdominal CT scans holds substantial potential to enhance disease diagnosis while reducing the time and effort needed for organ annotations. However, accurate segmentation of small and complex organs is still a challenge due to limited annotations. In multi-organ segmentation, the anatomical structure can provide rich semantic information such as relative position of organs. Therefore, we propose a text-guided cross pseudo supervision (TG-CPS) method, which integrates text embeddings from vision-language models (VLMs) into CPS framework. Specifically, we introduce a text branch to capture anatomical relationships using text label prompts, guiding the segmentation model to learn structured feature embeddings. Based on the text branch, unlabeled images can utilize anatomical relationships to generate more reliable pseudo labels. Furthermore, we design a multimodal fusion module to effectively integrate text and image features, thereby enhancing network learning for small and complex organs. Experiments on the FLARE2022 and AMOS datasets demonstrate that that the TG-CPS achieves more accurate multi-organ segmentation than the state-of-the-art semi-supervised segmentation methods, especially for small and complex organs.