Semi-supervised pulmonary airway segmentation based on dynamically decoupling intra-class regions

doi:10.11996/JG.j.2095-302X.2025040763

Abstract

Abstract:

Accurate airway segmentation from computed tomography (CT) images served as the foundation for diagnosing and treating various pulmonary diseases. However, the complex tree-like structure rendered pixel-level annotation extremely difficult. Semi-supervised learning methods provide insights for airway segmentation with limited labeled data. However, in airway segmentation tasks, there are significant intra-class differences between the large airway (trachea and main branches) and small airway (peripheral bronchioles) regarding voxel quantity, branch count, and morphological structure. Nevertheless, in airway segmentation tasks, the severe intra-class imbalance problem causes the model being prone to overfitting toward the dominated segmentation classes in semi-supervised learning with limited annotations. This results in insufficient representation learning for peripheral bronchioles, leading to poor segmentation accuracy and limiting clinical applications. To address this issue, a novel semi-supervised pulmonary airway segmentation framework based on a single-teacher dual-student three-branch network was proposed, achieving accurate airway tree-like structure with limited labeled data. In addition, a plug-and-play dynamic threshold module was developed to guide the network in identifying sub-regions with different segmentation difficulties during network training steps. Moreover, a novel intra-class region decoupling strategy was designed, enabling representation learning for sub-regions with different segmentation difficulties through different constraint optimization methods. Experimental results on two public datasets and a real-world dataset demonstrated that the proposed method outperformed the current state-of-the-art methods for airway segmentation. The Dice coefficient achieved 91.96%, while the TD and BD metrics reached 81.88% and 78.32%, respectively, enabling fast and accurate airway segmentation from CT images.

Key words: deep learning, airway segmentation, semi-supervised learning, decoupling intra-class regions, dynamic threshold

CLC Number:

WANG Ziyu, CAO Weiwei, CAO Yuzhu, LIU Meng, CHEN Jun, LIU Zhaobang, ZHENG Jian. Semi-supervised pulmonary airway segmentation based on dynamically decoupling intra-class regions[J]. Journal of Graphics, 2025, 46(4): 763-774.

Figures/Tables 12

References 32

[1]	尤堃, 郝鹏翼, 吴福理, 等. 基于三维卷积神经网络的肺结节假阳性筛查[J]. 图学学报, 2019, 40(3): 423-428. DOI
	YOU K, HAO P Y, WU F L, et al. False positive reduction of pulmonary nodules using 3D CNN[J]. Journal of Graphics, 2019, 40(3): 423-428 (in Chinese). DOI
[2]	蒋武君, 支力佳, 张少敏, 等. 基于通道残差嵌套U结构的CT影像肺结节分割方法[J]. 图学学报, 2023, 44(5): 879-889. DOI
	JIANG W J, ZHI L J, ZHANG S M, et al. CT image segmentation of lung nodules based on channel residual nested U structure[J]. Journal of Graphics, 2023, 44(5): 879-889 (in Chinese).
[3]	FETITA C I, PRETEUX F, BEIGELMAN-AUBRY C, et al. Pulmonary airways: 3-D reconstruction from multislice CT and clinical investigation[J]. IEEE Transactions on Medical Imaging, 2004, 23(11): 1353-1364. PMID
[4]	HOWLING S J, EVANS T W, HANSELL D M. The significance of bronchial dilatation on CT in patients with adult respiratory distress syndrome[J]. Clinical Radiology, 1998, 53(2): 105-109. DOI PMID
[5]	QIN Y L, ZHENG H, GU Y, et al. Learning tubule-sensitive CNNs for pulmonary airway and artery-vein segmentation in CT[J]. IEEE Transactions on Medical Imaging, 2021, 40(6): 1603-1617.
[6]	WANG P Y, GUO D Z, ZHENG D D, et al. Accurate airway tree segmentation in CT scans via anatomy-aware multi-class segmentation and topology-guided iterative learning[J]. IEEE Transactions on Medical Imaging, 2024, 43(12): 4294-4306.
[7]	CHARBONNIER J P, VAN RIKXOORT E M, SETIO A A A, et al. Improving airway segmentation in computed tomography using leak detection with convolutional networks[J]. Medical Image Analysis, 2017, 36: 52-60. DOI PMID
[8]	YUN J, PARK J, YU D, et al. Improvement of fully automated airway segmentation on volumetric computed tomographic images using a 2.5 dimensional convolutional neural net[J]. Medical Image Analysis, 2019, 51: 13-20. DOI PMID
[9]	ZHANG M H, WU Y Q, ZHANG H X, et al. Multi-site, multi-domain airway tree modeling[J]. Medical Image Analysis, 2023, 90: 102957.
[10]	QIN Y L, CHEN M J, ZHENG H, et al. AirwayNet: a voxel-connectivity aware approach for accurate airway segmentation using convolutional neural networks[C]// The 22nd International Conference on Medical Image Computing and Computer Assisted Intervention - MICCAI 2019. Cham: Springer, 2019: 212-220.
[11]	OUALI Y, HUDELOT C, TAMI M. Semi-supervised semantic segmentation with cross-consistency training[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 12671-12681.
[12]	CHEN X K, YUAN Y H, ZENG G, et al. Semi-supervised semantic segmentation with cross pseudo supervision[C]// 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2021: 2613-2622.
[13]	TARVAINEN A, VALPOLA H. Mean teachers are better role models: weight-averaged consistency targets improve semi- supervised deep learning results[C]// The 31st International Conference on Neural Information Processing Systems. New York: ACM, 2017: 1195-1204.
[14]	SHEN N, XU T F, BIAN Z Y, et al. SCANet: a unified semi-supervised learning framework for vessel segmentation[J]. IEEE Transactions on Medical Imaging, 2023, 42(9): 2476-2489.
[15]	YU L Q, WANG S J, LI X M, et al. Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation[C]// The 22nd International Conference on Medical Image Computing and Computer Assisted Intervention - MICCAI 2019. Cham: Springer, 2019: 605-613.
[16]	LUO X D, LIAO W J, CHEN J N, et al. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency[C]// The 24th International Conference on Medical Image Computing and Computer Assisted Intervention - MICCAI 2021. Cham: Springer, 2021: 318-329.
[17]	VERMA V, KAWAGUCHI K, LAMB A, et al. Interpolation consistency training for semi-supervised learning[J]. Neural Networks, 2022, 145: 90-106.
[18]	BAI Y H, CHEN D W, LI Q L, et al. Bidirectional copy-paste for semi-supervised medical image segmentation[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2023: 11514-11524.
[19]	BASAK H, GHOSAL S, SARKAR R. Addressing class imbalance in semi-supervised image segmentation: a study on cardiac MRI[C]// The 25th International Conference on Medical Image Computing and Computer Assisted Intervention - MICCAI 2022. Cham: Springer, 2022: 224-233.
[20]	LIN Y Q, YAO H F, LI Z Z, et al. Calibrating label distribution for class-imbalanced barely-supervised knee segmentation[C]// The 25th International Conference on Medical Image Computing and Computer Assisted Intervention - MICCAI 2022. Cham: Springer, 2022: 109-118.
[21]	ZHAO X Y, QI Z X, WANG S, et al. RCPS: rectified contrastive pseudo supervision for semi-supervised medical image segmentation[J]. IEEE Journal of Biomedical and Health Informatics, 2024, 28(1): 251-261.
[22]	SHEN Z Q, CAO P, YANG H, et al. Co-training with high-confidence pseudo labels for semi-supervised medical image segmentation[C]// The 32nd International Joint Conference on Artificial Intelligence. New York: ACM, 2023: 4199-4207.
[23]	XU Z, WANG Y X, LU D H, et al. Ambiguity-selective consistency regularization for mean-teacher semi-supervised medical image segmentation[J]. Medical Image Analysis, 2023, 88: 102880.
[24]	ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S S, et al.3D U-Net: learning dense volumetric segmentation from sparse annotation[C]// The 19th International Conference on Medical Image Computing and Computer-Assisted Intervention - MICCAI 2016. Cham: Springer, 2016: 424-432.
[25]	NORTHCUTT C, JIANG L, CHUANG I. Confident learning: estimating uncertainty in dataset labels[J]. Journal of Artificial Intelligence Research, 2021, 70: 1373-1411.
[26]	HE K M, ZHANG X Y, REN S Q, et al. Delving deep into rectifiers: surpassing human-level performance on ImageNet classification[C]// 2015 IEEE International Conference on Computer Vision. New York: IEEE Press, 2015: 1026-1034.
[27]	ARMATO S G, MCLENNAN G, BIDAUT L, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans[J]. Medical Physics, 2011, 38(2): 915-931. PMID
[28]	HOFMANNINGER J, PRAYER F, PAN J, et al. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem[J]. European Radiology Experimental, 2020, 4(1): 50. DOI PMID
[29]	ZHANG M H, GU Y. Towards connectivity-aware pulmonary airway segmentation[J]. IEEE Journal of Biomedical and Health Informatics, 2024, 28(1): 321-332.
[30]	LEE T C, KASHYAP R L, CHU C N. Building skeleton models via 3-D medial surface axis thinning algorithms[J]. CVGIP: Graphical Models and Image Processing, 1994, 56(6): 462-478.
[31]	CZOLBE S, ARNAVAZ K, KRAUSE O, et al. Is segmentation uncertainty useful?[C]// The 27th International Conference on Information Processing in Medical Imaging. Cham: Springer, 2021: 715-726.
[32]	HAN J F, LUO P, WANG X G. Deep self-learning from noisy labels[C]// 2019 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 5137-5146.

方法	标注使用比例		评价指标
方法	有标签	无标签	Dice	TD	BD
3D U-Net	10	0	85.76	67.64	56.18
3D U-Net	20	0	90.65	75.67	65.60
3D U-Net	100	0	93.48	84.38	76.48
MT	10	90	90.77	70.25	58.05
UA-MT	10	90	91.25	75.04	64.29
CPS	10	90	90.53	75.60	65.06
ICT	10	90	91.14	75.95	66.16
URPC	10	90	91.24	76.94	65.78
BCP	10	90	89.10	70.52	59.93
Ours	10	90	91.61	81.34	72.93
MT	20	80	91.46	77.47	67.28
UA-MT	20	80	91.49	77.10	67.56
CPS	20	80	91.41	78.87	69.57
ICT	20	80	91.12	79.89	70.90
URPC	20	80	90.74	78.93	69.01
BCP	20	80	90.55	71.84	61.10
Ours	20	80	91.96	81.88	73.82

方法	标注使用比例		评价指标
方法	有标签	无标签	Dice	TD	BD
3D U-Net	10	0	85.76	67.64	56.18
3D U-Net	20	0	90.65	75.67	65.60
3D U-Net	100	0	93.48	84.38	76.48
MT	10	90	90.77	70.25	58.05
UA-MT	10	90	91.25	75.04	64.29
CPS	10	90	90.53	75.60	65.06
ICT	10	90	91.14	75.95	66.16
URPC	10	90	91.24	76.94	65.78
BCP	10	90	89.10	70.52	59.93
Ours	10	90	91.61	81.34	72.93
MT	20	80	91.46	77.47	67.28
UA-MT	20	80	91.49	77.10	67.56
CPS	20	80	91.41	78.87	69.57
ICT	20	80	91.12	79.89	70.90
URPC	20	80	90.74	78.93	69.01
BCP	20	80	90.55	71.84	61.10
Ours	20	80	91.96	81.88	73.82

方法	标注使用比例		评价指标
方法	有标签	无标签	Dice	TD	BD
3D U-Net	10	0	86.77	59.87	48.32
3D U-Net	20	0	90.97	66.25	54.40
3D U-Net	100	0	92.30	79.89	71.43
MT	10	90	90.99	66.76	55.58
UA-MT	10	90	90.53	67.45	56.49
CPS	10	90	91.27	69.25	58.34
ICT	10	90	91.64	71.33	60.53
URPC	10	90	91.51	71.56	61.52
BCP	10	90	87.99	63.34	52.09
Ours	10	90	91.83	73.36	63.01
MT	20	80	91.66	70.17	59.83
UA-MT	20	80	91.60	71.46	61.01
CPS	20	80	92.24	73.17	62.41
ICT	20	80	92.27	74.56	64.80
URPC	20	80	91.63	73.59	64.46
BCP	20	80	91.05	70.34	59.54
Ours	20	80	92.34	75.80	66.02

方法	标注使用比例		评价指标
方法	有标签	无标签	Dice	TD	BD
3D U-Net	10	0	86.77	59.87	48.32
3D U-Net	20	0	90.97	66.25	54.40
3D U-Net	100	0	92.30	79.89	71.43
MT	10	90	90.99	66.76	55.58
UA-MT	10	90	90.53	67.45	56.49
CPS	10	90	91.27	69.25	58.34
ICT	10	90	91.64	71.33	60.53
URPC	10	90	91.51	71.56	61.52
BCP	10	90	87.99	63.34	52.09
Ours	10	90	91.83	73.36	63.01
MT	20	80	91.66	70.17	59.83
UA-MT	20	80	91.60	71.46	61.01
CPS	20	80	92.24	73.17	62.41
ICT	20	80	92.27	74.56	64.80
URPC	20	80	91.63	73.59	64.46
BCP	20	80	91.05	70.34	59.54
Ours	20	80	92.34	75.80	66.02

方法	标注使用比例		评价指标
方法	有标签	无标签	Dice	TD	BD
3D U-Net	10	0	80.61	49.22	40.37
3D U-Net	20	0	81.42	52.77	42.85
3D U-Net	100	0	83.78	60.56	48.99
MT	10	90	82.78	49.19	40.46
UA-MT	10	90	83.11	53.81	44.53
CPS	10	90	82.51	54.23	43.92
ICT	10	90	82.96	55.56	45.31
URPC	10	90	83.92	55.44	45.81
BCP	10	90	82.80	52.12	42.59
Ours	10	90	83.87	58.98	48.63
MT	20	80	83.64	56.95	46.56
UA-MT	20	80	83.40	57.22	46.73
CPS	20	80	82.78	55.81	45.61
ICT	20	80	83.63	58.80	47.92
URPC	20	80	83.57	58.43	47.91
BCP	20	80	82.84	52.53	43.32
Ours	20	80	83.88	59.40	48.54