KMS-Net: Kolmogorov–Arnold-Based Multi-Scale Attention Network for Cardiac Segmentation

Accurate segmentation of cardiac structures in 2D echocardiography is crucial for identifying heart disease and measuring functional metrics like chamber sizes and ejection fraction. Conventional U-Net models are good in extracting local features but face difficulties when dealing with long-range dependencies in noisy ultrasound images, whereas pure Transformers capture global context at the cost of fine boundary detail.We propose KMT-Net, a novel hybrid architecture that effectively combines Kolmogorov–Arnold Networks (KANs) with modern multi-scale attention mechanisms. Spline-based KAN layers (grid size 7) provide highly expressive nonlinear modeling, residual convolutional blocks incorporate Efficient Multi-Scale Attention (EMA), an enhanced KAN-based Atrous Spatial Pyramid Pooling with Squeeze-and-Excitation (KASPPS) aggregates rich multi-scale context, Selective State-Space (SS2D) layers enable efficient long-range spatial modeling, and Multi-Scale Attention Gates (MSAG) refine skip connections for precise boundary delin- eation.On the CAMUS dataset (500 patients, 2CH/4CH views at ED/ES), KMT-Net achieves mean Dice scores of 0.9417 and 0.9241 (2CH) and 0.9518 and 0.9328 (4CH), clearly outperforming U-Net variants, TransUNet, Swin-UNet, U-Mamba, and recent KAN-based models. It attains state-of-the-art performance on the ACDC MRI dataset (mean Dice 92.65%, HD95 1.08 mm) and exhibits strong zero-shot generalization to the EchoNet-Dynamic echocardiography dataset, confirming its capability to robustly manage both local details and global context across various imaging modalities and clinical environments.The code of KMS-Net is publicly available at: https://github.com/Hassan48khan/ KMS-Net-Kolmogorov-Arnold-Based-Multi-Scale-Attention-Network-for-Cardiac-Image-Segmentation.