Echo-Mamba: A Lightweight Mamba-Based Framework for Generalizable Left Ventricular Segmentation Across Multiple...

Left ventricle (LV) segmentation in echocardiography is crucial for the assessment of cardiac function and the diagnosis of cardiovascular disease. In clinical practice, there is an increasing demand for accurate and lightweight segmentation solutions that can be deployed on small and low-cost medical imaging devices. Currently, most models are designed using Convolutional Neural Networks (CNNs) and Transformer methods. However, CNNs exhibit limitations in long-range modeling capabilities, while Transformers are constrained by their quadratic computational complexity. These shortcomings are unfavorable for left ventricle segmentation tasks, which require both strong global semantic information modeling for anatomical structures and lightweight design to meet clinical deployment needs on resource-constrained devices. We propose Echo-Mamba, a lightweight architecture that models long-range dependencies with linear computational complexity for left ventricular echocardiography segmentation. In addition, our proposed Adaptive Feature Enhancement (AFE) module dynamically adapts to inter-patient variations in cardiac structural dimensions and morphological changes. We compared the segmentation performance of Echo-Mamba against other state of-the-art methods on four publicly available echocardiography datasets: EchoNet-Dynamic, EchoNet-Pediatric, CAMUS, and HMC-QU. Echo-Mamba demonstrates generalizable segmentation performance across anatomical targets, age groups, and multiview echocardiography, while maintaining a lightweight architecture (0.323M parameters) and strong accuracy. The implementation of our approach is publicly available at https://github.com/Beliefs666/EchoMamba.