Search

A Lightweight Hybrid CNN-Transformer with Boundary-Aware Loss for Cardiac Cine-MRI Multi-Structure Segmentation

Author(s)

Yali Wan1,*, Deyi Xiong2

Affiliation(s)

1School of Information Technology and Engineering, Guangzhou College of Commerce, Guangzhou, China 2Chongqing Academy of Metrology and Quality Inspection, Chongqing, China *Corresponding Author

Abstract

Automated quantification of cardiac function depends on accurate delineation of the right-ventricular cavity (RV), the left-ventricular myocardium (MYO), and the left-ventricular cavity (LV) in cine cardiac MRI. We present a lightweight hybrid CNN-Transformer for the Automated Cardiac Diagnosis Challenge (ACDC) dataset, trained with a compound boundary-aware loss. The network couples a four-level convolutional encoder-decoder with a single self-attention block placed at the lowest-resolution bottleneck, which supplies global context while keeping inference cost close to that of a standard U-Net. Skip connections pass through a learnable refinement that emphasises boundary pixels, and the loss combines Dice, cross-entropy, and a distance-weighted boundary cross-entropy term. Under 3-fold cross-validation on ACDC the method reaches a mean foreground Dice of 0.859 ± 0.004, on par with the strongest baseline (SwinUNet-2D, 0.866 at fold 0), while giving the lowest boundary error on the left-ventricular cavity and the myocardium (LV 95th-percentile Hausdorff Distance (HD95) 1.33 ± 0.14 versus 1.63 for SwinUNet-2D at fold 0, an 18% reduction; MYO HD95 1.94 ± 0.05 versus 1.95, a marginal reduction) and remaining within 0.07 pixels of SwinUNet-2D on the RV boundary (RV HD95 3.03 ± 0.71 versus 3.10). An ablation on fold 0 identifies the boundary-aware skip refinement as the component with the largest effect on boundary localisation: removing it raises MYO HD95 from 1.98 to 3.02 (a 52.5% increase) and LV HD95 from 1.48 to 2.86 (a 93% increase) under the per-slice ED+ES protocol, while mDice rises from 0.863 to 0.911 — i.e. the refinement trades bulk overlap for boundary precision. The model has 9.95 M parameters and runs on a single 8 GB GPU. The results suggest that combining boundary-aware architectural design with a boundary-weighted loss is a useful direction for cine-MRI segmentation when labelled data are limited.

Keywords

Cardiac Cine-MRI; ACDC; CNN-Transformer Hybrid; Boundary-Aware Loss; Medical Image Segmentation

References

[1] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015, LNCS 9351:234-241. DOI:10.1007/978-3-319-24574-4_28. [2] Bernard O, Lalande A, Zotti C, et al. Deep Learning Techniques for Automatic MRI Cardiac Multi-structures Segmentation and Diagnosis: Is the Problem Solved. IEEE Transactions on Medical Imaging, 2018, 37(11): 2514-2525. DOI:10.1109/TMI.2018.2837502. [3] Oktay O, Schlemper J, Folgoc L L, et al. Attention U-Net: Learning Where to Look for the Pancreas. arXiv: 1804.03999, 2018. DOI:10.48550/arXiv.1804.03999. [4] Gu Z, Cheng J, Fu H, et al. CE-Net: Context Encoder Network for 2D Medical Image Segmentation. IEEE Transactions on Medical Imaging, 2019, 38(7):2281-2292. DOI:10.1109/TMI.2019.2903562. [5] Isensee F, Jaeger P F, Kohl S A A, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 2021, 18:203-211. DOI:10.1038/s41592-020-01008-z. [6] Chen J, Lu Y, Yu Q, et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation. arXiv: 2102.04306, 2021. DOI:10.48550/arXiv.2102.04306. [7] Cao H, Wang Y, Chen J, et al. Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation. European Conference on Computer Vision Workshops (ECCVW), 2022, LNCS 13803:205-218. DOI:10.1007/978-3-031-25066-8_9. [8] Hatamizadeh A, Tang Y, Nath V, et al. UNETR: Transformers for 3D Medical Image Segmentation. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022:1748-1758. DOI:10.1109/WACV51458.2022.00181. [9] Wang H, Xie S, Lin L, et al. Mixed Transformer U-Net For Medical Image Segmentation. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022:2390-2394. DOI:10.1109/ICASSP43922.2022.9746172. [10]Wang F, Wang B. Boundary-guided feature integration network with hierarchical transformer for medical image segmentation. Multimedia Tools and Applications, 2024, 83:8955-8969. DOI:10.1007/s11042-023-15948-z. [11]Ma J, Li F, Wang B. U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation. arXiv:2401.04722, 2024. DOI:10.48550/arXiv.2401.04722. [12]Ma J, Chen J, Ng M, et al. Loss odyssey in medical image segmentation. Medical Image Analysis, 2021, 71:102035. DOI:10.1016/j.media.2021.102035. [13]Kervadec H, Bouchtiba J, Desrosiers C, et al. Boundary loss for highly unbalanced segmentation. Medical Imaging with Deep Learning (MIDL), 2019, PMLR 102:285-296. [14]Karimi D, Salcudean S E. Reducing the Hausdorff Distance in Medical Image Segmentation With Convolutional Neural Networks. IEEE Transactions on Medical Imaging, 2020, 39(2):499-513. DOI:10.1109/TMI.2019.2930068. [15]Milletari F, Navab N, Ahmadi S A. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation Fourth International Conference on 3D Vision (3DV), 2016:565-571. DOI:10.1109/3DV.2016.79. [16]Taghanaki S A, Zheng Y, Zhou S K, et al. Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation. arXiv:1805.02798, 2018. DOI:10.48550/arXiv.1805.02798