Semi-Supervised Learning With Kolmogorov-Arnold Network for MRI Cardiac Segmentation

MRI cardiac segmentation plays a vital role for the diagnosis of cardiovascular disease. Recently, many studies have developed semi-supervised learning (SSL) algorithms for this purpose. However, two challenges are still unsolved, i.e., linear pattern modeling and limited perturbation space. To this end, we develop KS-Net, an innovative framework for MRI cardiac segmentation, incorporating the Kolmogorov-Arnold network (KAN) module and an SSL-based perturbation strategy. Specifically, this work designs a U-shaped network with the KAN module to be compatible with nonlinear high-level features. In addition, we introduce a dual-stream perturbation approach to investigate the predefined perturbation space at the image level and leverage SSL for discriminative representations. Our proposed KS-Net is trained and tested on MyoPS 2020 and automated cardiac diagnosis challenge (ACDC) datasets. Experimental results indicated that it effectively outperforms existing state-of-the-art (SOTA) methods in MRI cardiac segmentation.