Outlet Title

IEEE Access

Document Type

Article

Publication Date

3-16-2026

Abstract

Accurate segmentation of cardiac structures in 2D echocardiography is essential for diagnosing cardiovascular disease and computing clinical metrics such as chamber volumes and ejection fraction. Conventional U-Net architectures excel at extracting local spatial features but struggle with long-range dependencies inherent in noisy ultrasound images, while pure Transformer-based models capture global context at the expense of fine boundary detail. To address these limitations, we propose KMS-Net, a novel hybrid segmentation architecture that integrates Kolmogorov–Arnold Networks (KANs), a class of learnable, spline-based function approximators that replace fixed activation functions with trainable nonlinear mappings, alongside multi-scale attention mechanisms. Specifically, spline-based KAN layers (grid size of 7) provide highly expressive nonlinear feature modeling. Residual convolutional blocks incorporate Efficient Multi-Scale Attention (EMA) and a KAN-based Atrous Spatial Pyramid Pooling with Squeeze-and-Excitation module (KASPPS) that aggregates rich multi-scale contextual information. Selective State-Space (SS2D) layers enable computationally efficient long-range spatial dependency modeling, and Multi-Scale Attention Gates (MSAG) refine skip connections for precise boundary delineation. Evaluated on the CAMUS dataset (500 patients, two-chamber/four-chamber views at end-diastole/end-systole), KMS-Net achieves mean Dice scores of 0.9417 and 0.9241 (2-CH view) and 0.9518 and 0.9328 (4-CH view), outperforming U-Net variants, TransUNet, Swin-UNet, U-Mamba, and recent KAN-based models. The proposed method further attains state-of-the-art performance on the ACDC cardiac MRI benchmark (mean Dice: 92.65%, HD95: 1.08 mm) and demonstrates strong zero-shot generalization to the EchoNet-Dynamic dataset, confirming its robustness across diverse imaging modalities and clinical settings.

Download

Share

COinS