A semi-supervised MRI image segmentation dual-network model for regions with ambiguous boundaries and heterogeneous regions

doi:10.12122/j.issn.1673-4254.2026.08.23

Abstract

Abstract:

Objective To construct a high-value region-guided dual-network semi-supervised segmentation method (HVASS) under extremely low labeling rates for addressing the challenges of heterogeneity and blurred boundaries in MRI tumor subregions and improving the accuracy and reliability of complex boundary segmentation for the heart and glioma. Methods HVASS employs a dual-network collaborative learning architecture that leverages prediction inconsistency to automatically identify two clinically significant high-risk regions: the high-confidence ambiguity zones, where predictions between the two networks diverge despite at least one network exhibiting high confidence; and the low-confidence stable zones, where predictions agree but with consistently low confidence across both networks. To refine pseudo-label quality, an adaptive dual-teacher mutual distillation mechanism was introduced for dynamically leveraging complementary knowledge from both networks. A high-value region-aware convolutional module was integrated to strengthen feature representation at the tumor margins and the heterogeneous areas, while a wavelet-based frequency-domain refinement module was incorporated to preserve the fine-grained edge details. The framework was evaluated on two publicly available datasets: BraTS2019 for brain glioma MRI and ACDC for cardiac MRI. Results On the ACDC dataset, HVASS achieved a mean Dice score of 90.34% and a 95th percentile Hausdorff Distance (95HD) of 2.46 mm, representing a 3.24% improvement in Dice and a 2.68 mm reduction in 95HD compared to the state-of-the-art model. On BraTS2019, the model attained a Dice score of 85.07% and a 95HD of 7.68 mm, with a 3.8% increase in Dice for enhancing tumor sub-region and a substantial reduction of boundary localization error. Conclusion HVASS demonstrates superior segmentation performance under minimal annotation settings and allows effective capture of fuzzy boundaries and heterogeneous tumor regions in MRI. The method shows particular strength in segmenting small lesions and ill-defined edges and thus lessens the annotation burden of the radiologists.

Key words: cardiac magnetic resonance imaging, glioma, image segmentation, semi-supervised learning, deep learning, fuzzy boundary

Lingxiao HUANG, Haizhe XU, Lingyan HUANG, Xinbo YAO, Kaiyuan ZHOU, Yongzhan GAO. A semi-supervised MRI image segmentation dual-network model for regions with ambiguous boundaries and heterogeneous regions[J]. Journal of Southern Medical University, 2026, 46(8): 1956-1966.

Figures/Tables 13

Fig.1 Sample two-dimensional slice images of the ACDC dataset. A: Transverse section slice of the base of the heart. B: Transverse section tissue slice of the middle part of the heart. C: Transverse section tissue slice of the middle part of the ventricular wall of the heart. D: Transverse section tissue slice of the distal part of the heart. E: Transverse section tissue slice of the apex of the heart.

Fig.2 Sample 2D slice images of the BraTS2019 dataset. A: Axial horizontal cross-section of the brain. B: Parasagittal sagittal section of the brain. C: Coronal section of the brain. D: Axial horizontal cross-section of the brain.

Fig.3 Distribution of image samples in ACDC and BraTS2019 datasets.

Fig.4 Heatmap of double-network pixel-level confidence type spatial distribution.

Fig.5 HVASS model network architecture.

Fig.6 ADTMG module structure diagram.

Fig.7 HVAC module structure diagram.

Fig.8 WFFR module structure diagram.

Tab.1 Experimental environment configuration

Configuration name	Model version
System environment	Ubuntu20.04
Central processing unit	Intel(R) Xeon(R) Gold 5418Y
Graphics processing unit	NVIDIA GeForce RTX 3090 24GB
Graphics processing unit acceleration library	CUDA 12.2
Random access memory	32 GB
Deep learning framework	Pytorch 2.0.0

Fig.9 Visualization results of segmentation by different algorithms. A: Original 2D axial MRI image (A1: Two-dimensional transverse section tissue slice of the middle part of the ventricular wall of the heart. A2: Two-dimensional transverse section tissue slice of the distal part of the heart. A3: Two-dimensional transverse section slice of the base of the heart. A4: Parasagittal sagittal section of the brain. A5: Axial horizontal cross-section of the brain. A6: Coronal section of the brain.). B: Ground truth annotation map. C: Lesion localization heatmap generated by the EVIL method. D: Lesion localization heatmap generated by the HAVSS method proposed in this study.

Tab.2 Comparison results of HVASS on the ACDC dataset

Method	Labeled	Unlabeled	Dice	IoU	95HD	ASD
MT_NIPS'2017	7 (10%)	63 (90%)	83.65	73.15	13.45	3.61
MCF_CVPR'2023			85.18	75.22	10.73	2.78
CPS_CVPR'2021			85.34	75.50	8.78	2.37
MC-Net_MICCAI'2021			86.07	76.58	11.48	3.37
AC-MT_MIA'2023			86.36	76.86	9.64	2.59
AAU_MIA'2024			86.46	76.99	8.00	2.06
UG-MCL_AIIM'2023			86.84	77.60	7.77	1.94
MLRP_MIA'2024			87.10	78.16	4.91	1.30
EVIL_CBM'2024			87.25	78.18	5.31	1.35
Ours-HVASS			90.34	83.13	2.46	0.79
MT_NIPS'2017	14 (20%)	56 (80%)	86.55	77.49	6.86	2.15
MC-Net_MICCAI'2021			86.58	77.68	15.99	4.93
CPS_CVPR'2021			87.03	78.13	6.66	2.12
MCF_CVPR'2023			87.32	78.43	6.53	2.00
AAU_MIA'2024			87.71	78.92	8.51	2.14
AC-MT_MIA'2023			87.99	79.34	7.85	2.14
MLRP_MIA'2024			88.00	79.54	5.41	1.58
EVIL_CBM'2024			88.34	79.90	6.46	1.68
UG-MCL_AIIM'2023			88.40	79.86	9.12	2.44
Ours-HVASS			91.02	83.65	2.74	0.78

Tab.3 Comparison results of HVASS on the BraTS2019 dataset

Method	Labeled	Unlabeled	Dice	IoU	95HD	ASD
MT_NIPS'2017	50 (20%)	200 (80%)	80.49	69.99	11.71	3.46
MCF_CVPR'2023			81.96	71.76	13.96	3.99
AC-MT_MIA'2023			82.09	71.65	15.91	4.81
AAU_MIA'2024			82.43	72.33	13.83	4.18
MC-Net_MICCAI'2021			82.51	72.98	14.57	3.78
EVIL_CBM'2024			82.60	72.46	11.44	2.76
CPS_CVPR'2021			82.70	72.30	12.99	3.46
UG-MCL_AIIM'2023			82.85	72.69	11.12	2.28
MLRP_MIA'2024			83.96	74.00	15.16	4.21
Ours-HVASS			85.07	75.80	7.68	2.03

Tab.4 Ablation experiment results of HVASS on the ACDC and BraTS2019 datasets under 10% annotation

Datasets	ADTMG	HVAC	WFFR	Dice	IoU	95HD	ASD
ACDC				84.38	71.45	18.23	4.61
	√			87.23	80.30	8.78	1.87
		√		86.97	79.28	11.48	3.37
			√	85.56	79.42	10.73	3.97
	√	√		88.78	80.22	3.86	0.98
	√		√	89.08	82.59	3.79	1.23
		√	√	88.89	81.43	5.31	1.58
	√	√	√	90.34	83.13	2.46	0.79

References 31

[1]	Litjens G, Ciompi F, Wolterink JM, et al. State-of-the-art deep learning in cardiovascular image analysis[J]. JACC Cardiovasc Imaging, 2019, 12(8 Pt 1): 1549-65. doi：10.1016/j.jcmg.2019.06.009
[2]	Bai WJ, Sinclair M, Tarroni G, et al. Automated cardiovascular magnetic resonance image analysis with fully convolutional networks[J]. J Cardiovasc Magn Reson, 2018, 20(1): 65. doi：10.1186/s12968-018-0471-x
[3]	周昊, 曾栋, 边兆英, 等. 基于半监督网络的组织感知CT图像对比度的增强方法[J]. 南方医科大学学报, 2023, 43(6): 985-93.
[4]	丁家满, 刘楠, 周蜀杰, 等. 基于正则化的半监督弱标签分类方法[J]. 计算机学报, 2022, 45(1): 69-81
[5]	翟德明, 沈斯娴, 周雄, 等. 基于最优传输理论的深度半监督学习伪标签生成算法[J]. 软件学报, 2024, 35(11): 5196-209.
[6]	Li CC, Zhang JS, Niu DM, et al. Boundary-aware uncertainty suppression for semi-supervised medical image segmentation[J]. IEEE Trans Artif Intell, 2024, 5(8): 4074-86. doi：10.1109/tai.2024.3359576
[7]	姚宗亮, 黄荣, 董爱华, 等. 基于多模态融合和自适应剪枝Transformer的脑肿瘤图像分割算法[J]. 宁夏大学学报: 自然科学版, 2024, 45(1): 16-24.
[8]	徐晗晗, 张印辉, 何自芬, 等. 伪标签置信度调控结直肠癌病理图像半监督语义分割[J]. 光学精密工程, 2025, 33(4): 591-609.
[9]	Chen XK, Yuan YH, Zeng G, et al. Semi-supervised semantic segmentation with cross pseudo supervision[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 20-25, 2021. Nashville, TN, USA. IEEE, 2021: 2613-22. doi：10.1109/cvpr46437.2021.00264
[10]	Wu YC, Xu MF, Ge ZY, et al. Semi-supervised left atrium segmentation with mutual ConsistencyTraining[C]//Medical Image Computing and Computer Assisted Intervention-MICCAI 2021. Cham: Springer, 2021: 297-306. doi：10.1007/978-3-030-87196-3_28
[11]	Adiga V S, Dolz J, Lombaert H. Anatomically-aware uncertainty for semi-supervised image segmentation[J]. Med Image Anal, 2024, 91: 103011. doi：10.1016/j.media.2023.103011
[12]	Bernard O, Lalande A, Zotti C, et al. Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?[J]. IEEE Trans Med Imaging, 2018, 37(11): 2514-25. doi：10.1109/tmi.2018.2837502
[13]	Bakas S, Akbari H, Sotiras A, et al. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features[J]. Sci Data, 2017, 4: 170117. doi：10.1038/sdata.2017.117
[14]	权笑, 戴鹏程, 辛潮, 等. LTE网络高价值区域识别与分析方法研究[J].电信工程技术与标准化, 2018, 31(5):18-21.
[15]	Zhou ZW, Rahman Siddiquee MM, Tajbakhsh N, et al. UNet++: a nested U-Net architecture for medical image segmentation[C]//Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Cham: Springer, 2018: 3-11. doi：10.1007/978-3-030-00889-5_1
[16]	Zhao CJ, Xiang S, Wang YQ, et al. Context-aware network fusing transformer and V-Net for semi-supervised segmentation of 3D left atrium[J]. Expert Syst Appl, 2023, 214: 119105. doi：10.1016/j.eswa.2022.119105
[17]	Topol EJ. High-performance medicine: the convergence of human and artificial intelligence[J]. Nat Med, 2019, 25(1): 44-56. doi：10.1038/s41591-018-0300-7
[18]	Tarvainen A, Valpola H. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results[C]//Neural Information Processing Systems., 2017
[19]	Xu Z, Wang YX, Lu DH, et al. Ambiguity-selective consistency regularization for mean-teacher semi-supervised medical image segmentation[J]. Med Image Anal, 2023, 88: 102880. doi：10.1016/j.media.2023.102880
[20]	Zhang YC, Jiao RS, Liao QC, et al. Uncertainty-guided mutual consistency learning for semi-supervised medical image segmentation[J]. Artif Intell Med, 2023, 138: 102476. doi：10.1016/j.artmed.2022.102476
[21]	Wang YC, Xiao B, Bi XL, et al. MCF: mutual correction framework for semi-supervised medical image segmentation[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 17-24, 2023, Vancouver, BC, Canada. IEEE, 2023: 15651-60. doi：10.1109/cvpr52729.2023.01502
[22]	Chen YY, Yang ZY, Shen CY, et al. Evidence-based uncertainty-aware semi-supervised medical image segmentation[J]. Comput Biol Med, 2024, 170: 108004. doi：10.1016/j.compbiomed.2024.108004
[23]	郑孙易, 刘佳鑫, 崔效楠, 等. 人工智能在肿瘤影像学中的应用进展[J]. 放射学实践, 2025, 40(9): 1093-7.
[24]	Menze BH, Jakab A, Bauer S, et al. The multimodal brain tumor image segmentation benchmark (BRATS)[J]. IEEE Trans Med Imaging, 2015, 34(10): 1993-2024.
[25]	Bauer S, Wiest R, Nolte LP, et al. A survey of MRI-based medical image analysis for brain tumor studies[J]. Phys Med Biol, 2013, 58(13): R97-R129. doi：10.1088/0031-9155/58/13/r97
[26]	Salvador A, Bellver M, Campos V, et al. Recurrent neural networks for semantic instance segmentation[J]. arXiv preprint arXiv:, 2017.
[27]	Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool[J]. BMC Med Imaging, 2015, 15: 29. doi：10.1186/s12880-015-0068-x
[28]	Petitjean C, Dacher JN. A review of segmentation methods in short axis cardiac MR images[J]. Med Image Anal, 2011, 15(2): 169-84. doi：10.1016/j.media.2010.12.004
[29]	Zhuang XH, Shen J. Multi-scale patch and multi-modality atlases for whole heart segmentation of MRI[J]. Med Image Anal, 2016, 31: 77-87. doi：10.1016/j.media.2016.02.006
[30]	Isensee F, Jaeger PF, Kohl SAA, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation[J]. Nat Methods, 2021, 18(2): 203-11. doi：10.1038/s41592-020-01008-z
[31]	Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare[J]. Nat Med, 2019, 25(1): 24-9. doi：10.1038/s41591-018-0316-z