Multi-Scale Feature Fusion and Visual State Space Models for Echocardiogram Segmentation

Zihe Luo¹Fufeng Wang¹Xinkui Liao²Wei Lv^1,3Xiaolin Zhu^1,3*

• ¹Zhuhai College of Science and Techology, Zhuhai, China
• ²School of Medicine, Jinan University, Guangzhou, Guangdong Province, China
• ³City University of Macau, Macao, Macao, SAR China

Echocardiography plays a vital role in early cardiac disease intervention, enabling accurate and noninvasive cardiac function assessment. However, echocardiogram segmentation remains challenging due to low image quality, anatomical variability, and the need to model both local structures and global contextual information. CNNs are effective at local feature extraction but lack global modeling capability, while Transformer-based models offer stronger context reasoning at the cost of quadratic complexity and heavy parameters. To address these limitations, we propose EchoUMamba, a multiscale feature fusion network built upon Mamba, a recently introduced state space model that enables global context modeling with linear complexity. By leveraging the structured state-space formulation, EchoUMamba reduces computational cost while maintaining long-range feature representation, offering a principled alternative to conventional self-attention mechanisms. EchoUMamba integrates a multi-scale visual state-space block (EchoVSSB) to facilitate cross-scale feature interaction and global information flow. Residual fusion pathways are employed to enhance low-level feature preservation and gradient propagation. Unlike prior Mamba-based designs trained from scratch, our model utilizes ImageNet pretraining to improve convergence and generalization in medical imaging. Extensive experiments on publicly available datasets validate the effectiveness of our approach. EchoUMamba achieves a precision of 93.75%, recall of 95.28%, Dice score of 94.51%, and IoU of 89.66%. It also exhibits strong efficiency, requiring only 26.3M parameters, 32.7 GFLOPs, and 22.8 ms inference time.