کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Mamba-SAM: A Hybrid Architecture for Efficient Cardiac MRI Medical Image Segmentation

نویسندگان :

Mohammadreza Gholipour Shahraki¹ Mehdi Rezaeian² Mohammad Ghasemzadeh³

1- دانشگاه یزد 2- دانشگاه یزد 3- دانشگاه یزد

کلمات کلیدی :

Medical Image Segmentation،Foundation Models،Segment Anything Model،State-Space Models،Mamba،Hybrid Architectures،Cardiac MRI

چکیده :

The Segment Anything Model (SAM) exhibits poor performance on medical images due to the domain gap from its natural image training data, high computational cost, and inherent 2D design. This paper introduces Mamba-SAM, a novel hybrid framework that efficiently adapts SAM by integrating it with a Visual Mamba (VMamba) encoder. Our architecture leverages a frozen SAM backbone for general feature extraction while a trainable VMamba branch captures domain-specific details. A Cross-Branch Attention module fuses these complementary features, and an Implicit Feature Alignment decoder ensures precise segmentation. On the ACDC cardiac MRI dataset, Mamba-SAM achieves a highly competitive average Dice score of 0.906, closely matching specialized models like UNet++, while being significantly more parameter-efficient. This work demonstrates that combining foundation models with modern state-space architectures is a powerful strategy for accurate and efficient medical image analysis.