Cascaded LGE-MRI Myocardial Fibrosis Segmentation With Dirichlet Output Parametrization and Discretization-Step Aware Regularization
DOI:
https://doi.org/10.32603/2071-2340-2026-2-43-56Keywords:
LGE-MRI, myocardial fibrosis, segmentation, Dirichlet distribution, uncertainty estimation, cascaded neural networks, normalized finite differencesAbstract
Introduction. Automatic myocardium and post-infarction scar segmentation in LGE-MRI is clinically important for quantitative assessment, yet it is sensitive to anisotropic acquisition and domain shift. The aim of this study is to evaluate whether a cascaded design with uncertainty-aware output and discretization-step-aware regularization improves segmentation accuracy and robustness. Methods. We propose FibroSegNet, a three-stage slice-wise cascade (CropNet, MyoSegNet, ScarSegNet) where ScarSegNet outputs are Dirichlet-parameterized and regularized by Rh, a finite-difference term normalized by physical grid spacing. The model is evaluated on EMIDEC (n=20) against 2D nnU-Net, 2D Swin UNETR, and 3D nnU-Net under the same protocol, with external validation on MyoPS (n=25), ablation studies, and anisotropic resampling tests. Results. On EMIDEC, FibroSegNet reached Dice 0.891/0.845 (myocardium/scar) versus 0.864/0.812 (2D nnU-Net), 0.868/0.819 (Swin UNETR), and 0.838/0.776 (3D nnU-Net); all myocardium Dice differences remained significant after Bonferroni correction (pBonf < 0.004). On MyoPS without retraining, myocardium performance remained better (+0.032, p < 0.001) while scar differences were not significant (p = 0.38); ablation indicated cascade dominance, Dirichlet benefits for calibration (ECE reduced by 0.026) and error detection (AUC ROC 0.979), and a 1.28 mm HD95 reduction from Rh. Discussion. The results show that cascade decomposition is the main contributor under strong LGE-MRI anisotropy, while uncertainty maps are primarily valuable for reliability control rather than overlap gain. The main limitation is cross-domain scar transferability, motivating target-domain fine-tuning and domain adaptation.
References
R. Karim, P. Bhagirath, P. Claus et al., "Evaluation of state-of-the-art segmentation algorithms for left ventricle infarct from late gadolinium enhancement MR images," Medical Image Analysis, vol. 30, pp. 95-107, 2016, doi: 10.1016/j.media.2016.01.004.
A. Lalande, Z. Chen, T. Pommier et al., "Deep learning methods for automatic evaluation of delayed enhancement MRI: the results of the EMIDEC challenge," Medical Image Analysis, vol. 79, p. 102428, 2022, doi: 10.1016/j.media.2022.102428.
F. Isensee, P. F. Jaeger, S. A. Kohl et al., "nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation," Nature Methods, vol. 18, no. 2, pp. 203-211, 2021, doi: 10.1038/s41592-020-01008-z.
M. Sensoy, L. Kaplan, and M. Kandemir, "Evidential deep learning to quantify classification uncertainty," in Advances in Neural Information Processing Systems, vol. 31. Curran Associates, Inc., 2018, pp. 3179-3189.
C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, "On calibration of modern neural networks," in Proc. of 34th International Conference on Machine Learning (PMLR), no. 70, 2017, pp. 1321-1330.
S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon, "CBAM: Convolutional block attention module," in European Conference on Computer Vision, Springer, 2018, pp. 3-19, doi: 10.1007/978-3-030-01234-2_1.
H. Kervadec, J. Bouchitba, C. Desrosiers et al., "Boundary loss for highly unbalanced segmentation," in Proc. of 2nd International Conference on Medical Imaging with Deep Learning (PMLR), no. 102, 2019, pp. 285-296.
A. Hatamizadeh, V. Nath, Y. Tang et al., "Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images," in Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, Springer, 2022, pp. 272-284, doi: 10.1007/978-3-031-08999-2_22.
H.-Y. Zhou, J. Guo, Y. Zhang et al., "nnFormer: Interleaved transformer for volumetric segmentation," IEEE Transactions on Image Processing, vol. 32, pp. 4036-4048, 2023, doi: 10.1109/TIP.2023.3293771.
A. Malinin and M. Gales, "Predictive uncertainty estimation via prior networks," in Advances in Neural Information Processing Systems, vol. 31. Curran Associates, Inc., 2018, pp. 7047-7058.
Y. Tan, F. Feng, and J. Zhao, "A benchmarking study of deep learning approaches for bi-arial segmentation on late gadolinium-enhanced MRIs," in Statistical Atlases and Computational Models of the Heart: STACOM 2023, in Lecture Notes in Computer Science. Springer, 2024, pp. 249-259, doi: 10.1007/978-3-031-52448-6_24.
M. Schwab, M. Pamminger, C. Kremser, M. Haltmeier, and A. Mayr, "Deep learning pipeline for fully automated myocardial infarct segmentation from clinical cardiac MR scans," Radiology Advances, vol. 2, no. 4, p. umaf023, 2025, doi: 10.1093/radadv/umaf023.
E. Yu. Shchetinin and A. A. Shevchuk, "Computational image segmentation method based on a Dirichlet field and asymptotic discretization accuracy analysis of spatial regularizers," Model Optim Inf Technol, vol. 14, no. 3, 2026, (in Russian), doi: 10.26102/2310-6018/2026.54.3.009.
L. Li, F. Wu, S. Wang et al., "MyoPS: a benchmark of myocardial pathology segmentation combining three-sequence cardiac magnetic resonance images," Medical Image Analysis, vol. 87, p. 102808, 2023, doi: 10.1016/j.media.2023.102808.
O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional networks for biomedical image segmentation," in Medical Image Computing and Computer-Assisted Intervention — MICCAI 2015, Springer, 2015, pp. 234-241, doi: 10.1007/978-3-319-24574-4_28.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Evgeniy Shchetinin, Vladislav Ruzmanov, Andrey Shevchuk
This work is licensed under a Creative Commons Attribution 4.0 International License.
This work is licensed under a Creative Commons Attribution 4.0 International License.
