Research on Myocardial and Cardiac Chamber Segmentation in Cardiac MR Images Based on DRMGU-Net

Ning Li¹Xianfeng Guo²Jiangshan Cao³Baolong Shi¹Bingli Liu⁴Yifei Wang⁴Baoqing Yu^4*Shangjun Huang^4*Zhu Shuxian^3*

• ¹School of Physical Science and Technology, Suzhou University of Science and Technology, Suzhou, China
• ²Department of Ultrasound Medicine, Seventh People's Hospital of Shanghai University of Traditional Chinese Medicine, Shanghai, China
• ³School of Electronics and Information Engineering, Suzhou University of Science and Technology, Suzhou, China
• ⁴Department of Orthopedics and Traumatology, Seventh People's Hospital of Shanghai University of Traditional Chinese Medicine, Shanghai, China

Obtaining precise cardiac magnetic resonance (MR) images is crucial for cardiovascular research. Conventional segmentation methods often yield substantial inaccuracies due to the heart's complex anatomy, significant background noise, and indistinct structural boundaries. To address these limitations, this study introduces a novel cardiac MR image segmentation network, DRMGU-Net, built upon the U-Net architecture. The encoder incorporates a dense convolutional block for efficient multi-dimensional feature extraction with reduced computational cost. Residual convolutional blocks are added to improve gradient flow and model stability. A multi-scale atrous spatial fusion attention module captures fine details and broadens the receptive field, enhancing multi-scale feature recognition. Skip connections integrate spatial and channel attention mechanisms to suppress non-target interference and balance local details with global context. A combined Dice and weighted cross-entropy loss function is employed, and the Mish activation function strengthens nonlinear representation. Comparative and ablation experiments on the ACDC dataset confirm that each proposed enhancement effectively boosts segmentation performance. DRMGU-Net achieves an average Dice coefficient of 92.6% and an average Hausdorff distance of 10.255 mm, demonstrating robust segmentation across five cardiac conditions. This work provides a reliable solution for automated cardiac MR image analysis and disease diagnosis.