Scholarlink

<p>With the widespread application of deep learning in medical imaging, U-Net and its transformer-based variants have achieved remarkable success. However, these models often suffer from high computational complexity and limitations in modeling local-global feature interactions, constraining their performance on intricate tasks like cardiac MRI segmentation. To address these challenges, we propose MCAN, a computationally efficient 2D segmentation framework based on UniFormer and multi-scale attention mechanisms. MCAN introduces three key modifications: (1) replacing the standard Swin-Transformer encoder with a UniFormer-based backbone to balance long-range dependency modeling with reduced computational overhead; (2) designing a Multi-scale Cascaded Fully Convolutional Attention Decoder to facilitate hierarchical feature fusion for enhancing complex anatomical boundaries; and (3) introducing an Efficient Multi-scale Convolutional Attention Module (MSCAM) and Large Kernel Grouped Attention Gates (LGAG) to adaptively focus on critical regions while suppressing background noise. Experimental results demonstrate that MCAN achieves a Mean DSC of 91.12% on ACDC and 78.34% on Synapse. Requiring only 26.45 M parameters and 9.32 G FLOPs, MCAN offers improved computational efficiency compared to larger models like SwinUNETR (384.20 G FLOPs). The ablation studies confirm the contribution of each module, indicating that MCAN provides a balanced and lightweight approach for clinical medical image analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MCAN: Cardiac MRI image segmentation method based on multi-scale convolution and attention mechanism

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Baoshan Sun,
  • Ziyan Weng,
  • Borui Liu

摘要

With the widespread application of deep learning in medical imaging, U-Net and its transformer-based variants have achieved remarkable success. However, these models often suffer from high computational complexity and limitations in modeling local-global feature interactions, constraining their performance on intricate tasks like cardiac MRI segmentation. To address these challenges, we propose MCAN, a computationally efficient 2D segmentation framework based on UniFormer and multi-scale attention mechanisms. MCAN introduces three key modifications: (1) replacing the standard Swin-Transformer encoder with a UniFormer-based backbone to balance long-range dependency modeling with reduced computational overhead; (2) designing a Multi-scale Cascaded Fully Convolutional Attention Decoder to facilitate hierarchical feature fusion for enhancing complex anatomical boundaries; and (3) introducing an Efficient Multi-scale Convolutional Attention Module (MSCAM) and Large Kernel Grouped Attention Gates (LGAG) to adaptively focus on critical regions while suppressing background noise. Experimental results demonstrate that MCAN achieves a Mean DSC of 91.12% on ACDC and 78.34% on Synapse. Requiring only 26.45 M parameters and 9.32 G FLOPs, MCAN offers improved computational efficiency compared to larger models like SwinUNETR (384.20 G FLOPs). The ablation studies confirm the contribution of each module, indicating that MCAN provides a balanced and lightweight approach for clinical medical image analysis.