Semi-supervised medical image segmentation using interpolation consistency training framework with UKAN and multi-task learning

In medical image analysis, accurate image segmentation is essential for the early diagnosis and treatment of diseases. However, traditional fully supervised learning methods often depend on extensive labeled data, which can be challenging to gather in the medical field due to high costs and time constraints. Therefore, there is an urgent need for an effective semi-supervised learning method that can fully utilize limited labeled data and abundant unlabeled data. We propose a semi-supervised medical image segmentation method based on the Interpolation Consistency Training (ICT) framework, termed UKMT-ICT. The method integrates a Kolmogorov-Arnold Network (KAN) and a multi-task learning strategy within a teacher–student architecture. The KAN module is introduced to construct UKAN, enhancing feature extraction and enabling more effective modeling of complex local and global information. In addition, a multi-task learning mechanism is employed to jointly predict segmentation masks (SM) and signed distance maps (SDM), facilitating information sharing between tasks and improving generalization and robustness. Experiments on the publicly available Automated Cardiac Diagnosis Challenge (ACDC) dataset demonstrate that UKMT-ICT outperforms six state-of-the-art semi-supervised segmentation methods. Using only 3% labeled data, our method achieves average improvements of 8.45% in Dice similarity coefficient (DSC) and 10.57% in Jaccard similarity coefficient (JSC), along with a 30.87% reduction in the 95% Hausdorff distance (HD95). The proposed UKMT-ICT method provides an effective solution for semi-supervised medical image segmentation under limited annotation settings, offering strong performance and practical potential for medical imaging applications.