动态小波多向感知与几何轴解引导三维CT骨折图像分割

张印辉; 刘凯; 何自芬; 张金凯; 陈光晨; 马志坚

doi:10.11999/JEIT250732

动态小波多向感知与几何轴解引导三维CT骨折图像分割

doi: 10.11999/JEIT250732 cstr: 32379.14.JEIT250732

1.
昆明理工大学，机电工程学院昆明 650500
2.
昆明市第一人民医院运动医学科昆明 650224

基金项目: 国家自然科学基金(62561032, 62171206、62061022)，云南省科技厅昆明医科大学联合专项《数字力学仿真优化踝关节骨折内固定》(202001AY070001-094)，昆明市卫健委十百千工程(2021-SW(后备)-28)，昆明市卫健委医学技术中心项目(2022-SW(技)-15)

详细信息

作者简介:
张印辉：男，博士，教授，研究方向为图像处理、机器视觉及机器智能

刘凯：男，硕士生，研究方向为医学图像处理

何自芬：女，博士，教授，研究方向为图像处理和机器视觉

张金凯：男，硕士生，研究方向为医学图像处理

陈光晨：男，博士生，研究方向为计算机视觉

马志坚：男，博士，副主任医师，研究方向为数字骨科

通讯作者:
马志坚　davidrochest@139.com

中图分类号: TN911.73; TP391.41
计量
- 文章访问数: 50
- HTML全文浏览量: 15
- PDF下载量: 5
- 被引次数: 0
出版历程
- 修回日期: 2025-12-12
- 录用日期: 2025-12-12
- 网络出版日期: 2025-12-19

Dynamic Wavelet Multi-Directional Perception and Geometry Axis-Solution Guided 3D CT Fracture Image Segmentation

1.
Mechanical and Electrical Engineering, Kunming University of Science and Technology, Kunming 650500, China
2.
Department of Sports Medicine, The First People’s Hospital of Kunming, Kunming 650224, China

Funds: The National Natural Science Foundation of China (62561032, 62171206, 62061022), Joint Special Project of Yunnan Provincial Department of Science and Technology and Kunming Medical University – ‘Digital Mechanics Simulation for Optimizing Internal Fixation of Ankle Fractures’ (202001AY070001-094), Kunming Municipal Health Commission Ten-Hundred-Thousand Talent Project (2021-SW(Reserve)-28), Kunming Municipal Health Commission Medical Technology Center Project (2022-SW(Tech)-15)

摘要

摘要: 三维骨折图像分割是临床骨科术前方案量化的关键，其中骨折面分割性能则直接影响手术决策的安全性与有效性。针对三维骨折图像分割中存在深层全局特征捕获不足、骨折面细节方向纹理模糊以及骨折图像几何结构利用不充分的问题，该文提出动态小波多向感知与几何轴解引导的三维CT骨折图像分割方法(Dynamic Wavelet Multi-Directional Perception and Geometry Axis-Solution Guided Network, DWAG-Net)。首先，为充分提取多维度视角下全局特征，设计动态可学习参数插值重构三平面视角特征，并与全维特征聚合实现多维度提取骨折图像全局信息。其次，引入三维小波变换，通过各方向高频子带的跨子带特征融合，增强模型对模糊骨折面中方向特征的纹理细节感知。最后，根据骨骼结构相似性与骨折面局部突变性设计几何轴解引导模块，通过几何亲和与距离衰减引导模型轴解分割，并重新分配类别权重缓解其不平衡问题，约束损失函数梯度向最优方向下降。在自建胫骨骨折数据集上，DWAG-Net模型相比现有先进模型展现出优越分割性能，平均Dice Score为71.20%其中骨折面类较基准提升了7.12%，平均HD95为1.38 mm其中骨折面类降低了3.70 mm，与前沿3D分割算法nnWNet相比平均Dice Score提升了4.63%。实验结果表明，DWAG-Net模型可有效实现三维骨折图像分割任务，为术前穿钉角度量化与固定位置精确选择提供了图像依据，对辅助医生优化术前方案具有重要意义。
- 三维骨折分割 /
- 胫骨CT图像 /
- 动态多视角 /
- 小波感知 /
- 几何引导
Abstract: Objective Accurate segmentation of fracture surfaces in 3D CT images is critical for orthopedic surgical planning, particularly in determining optimal nail insertion angles perpendicular to fracture planes. However, existing methods exhibit three key limitations: insufficient capture of deep global volumetric context, directional texture ambiguity in low-contrast fracture regions, and underutilization of geometric features decoding. To overcome these challenges, we propose DWAG-Net, a novel framework that integrates Dynamic Wavelet Perception and Geometry-Axis Guidance, to significantly enhance segmentation precision for complex tibial fractures and provide reliable 3D digital guidance for preoperative planning. Methods The architecture extends 3D nnU-Netv2 with three core innovations. First, the Dynamic Multi-View Aggregation (DMVA) module dynamically fuses tri-planar (axial/sagittal/coronal) and full-volume features via learnable parameter interpolation (optimized kernel size: 2×2×2) and channel-wise Hadamard product, enhancing global context modeling. Second, the Wavelet Direction Perception Enhancement (WDPE) module decomposes inputs using 3D Symlets discrete wavelet transform and applies direction-specific enhancement to eight subbands: adaptive convolutional kernels (e.g., [5,3,3] for depth-dominant fractures) amplify texture details in high-frequency subbands, while cross-subband fusion strategies integrate complementary features. Third, the Geometry Axis-Solution Guided (GASG) module embeds in decoders to enforce anatomical consistency by computing axis-level affinity maps (depth/height/width) that combine geometric similarity with spatial distance decay, and by optimizing boundary delineation through rotational positional encoding and multi-axis attention. The model was trained on the YN-TFS dataset (110 tibial fracture CT scans, resolution 0.39–1.00 mm) using SGD optimizer (lr=0.01, momentum=0.99) and a class-weighted loss (background:0.5, bone:1, fracture:5) to mitigate severe pixel imbalance. Results and Discussions DWAG-Net achieved state-of-the-art performance, with a mean Dice score of 71.20% (Table 1), surpassing nnU-Netv2 by 5.06% (fracture surface Dice: 69.48%, +7.12%). Boundary precision improved significantly, yielding a mean HD95 of 1.38 mm (fracture surface: 1.54 mm, –3.70 mm). Ablation studies (Table 2) confirmed each module’s contribution: DMVA improved Dice by 2.40% through adaptive multi-view fusion; WDPE resolved directional ambiguity, adding a 5.84% fracture-surface Dice gain; GASG provided a further 1.20% gain by enforcing geometric consistency. Key configurations included optimal DMVA parameters (2×2×2), Symlets wavelets, and sequential axis processing (D→H→W). Qualitatively, DWAG-Net preserved fracture integrity where U-Mamba/nnWNet failed and reduced over-segmentation compared with nnFormer/UNETR++ (Fig. 4). Conclusions DWAG-Net establishes a new state-of-the-art for 3D fracture segmentation by synergizing multi-directional wavelet perception and geometric guidance. Its key innovations—DMVA for volumetric context fusion, WDPE for directional texture enhancement, and GASG for anatomically consistent decoding—deliver clinically critical precision (71.20% Dice, 1.38 mm HD95) and enable data-driven surgical planning. Future work will focus on optimizing loss functions for severe class imbalance.
- 3D fracture segmentation /
- Tibial CT images /
- Dynamic multi-view /
- Wavelet perception /
- Geometric guidance

HTML全文

图 1 DWAG-Net模型总体框架

下载: 全尺寸图片幻灯片

图 2 动态多视角聚合模块

下载: 全尺寸图片幻灯片

图 3 小波方向感知模块

下载: 全尺寸图片幻灯片

图 4 几何轴解引导模块

下载: 全尺寸图片幻灯片

图 5 不同模型分割结果

下载: 全尺寸图片幻灯片

表 1 不同模型的分割性能对比

模型	年份	Dice Score/%			HD95/mm			FLOPs/G	Params/M	Inference/s
模型	年份	Avg	骨骼	骨折面	Avg	骨骼	骨折面	FLOPs/G	Params/M	Inference/s
nnU-Netv2	2023	66.14	69.93	62.36	3.68	2.12	5.24	561.92	30.79	124.00
nnFormer	2023	55.13	50.79	59.47	98.99	137.85	60.12	47.70	37.15	218.97
UNETR++	2024	51.13	46.42	55.84	103.39	137.97	68.80	43.70	66.80	213.56
U-Mamba	2024	68.86	71.38	66.35	5.10	6.19	4.02	1141.86	42.12	115.95
nnWNet	2025	66.57	66.98	66.15	1.94	1.77	2.11	1378.62	60.46	123.09
DWAG-Net	2025	71.20	72.91	69.48	1.38	1.22	1.54	675.22	57.60	159.48

下载: 导出CSV

表 2 所提出的模块对分割精度的影响

模型	模块			Dice Score /%			HD95 / mm
模型	DMVA	WDPE	GASG	Avg	骨骼	骨折面	Avg	骨骼	骨折面
基准				66.15	69.93	62.36	3.68	2.12	5.24
Ours¹	√			68.55	70.30	66.79	2.04	2.27	1.81
Ours²		√		67.50	67.21	67.79	1.84	1.92	1.75
Ours³	√	√		70.00	71.80	68.20	1.53	1.27	1.79
DWAG-Net	√	√	√	71.20	72.91	69.48	1.38	1.22	1.54

下载: 导出CSV

表 3 DMVA与WDPE编码位置对分割精度的影响

DMVA(√)与WDPE(O)嵌入位置			Dice Score /%			HD95 / mm
3	4	5	Avg	骨骼	骨折面	Avg	骨骼	骨折面
√ O	√ O		69.72	72.51	66.92	1.50	1.22	1.78
√ O	√	O	68.75	72.57	64.92	1.77	1.25	2.29
√	√ O	O	68.43	70.82	66.03	1.55	1.32	1.77
√ O	O	√	67.06	68.59	65.53	2.03	2.15	1.90
√ O		√ O	67.92	71.02	64.82	1.71	1.63	1.79
√	O	√ O	70.00	71.80	68.20	1.53	1.27	1.79
O	√ O	√	67.56	69.92	65.20	1.67	1.58	1.76
O	√	√ O	68.49	70.88	66.09	1.55	1.31	1.79
	√ O	√ O	67.70	69.40	66.00	2.04	2.30	1.78

下载: 导出CSV

表 4 DMVA模块中动态可学习参数大小对分割精度的影响

动态可学习参数大小	Dice Score /%			HD95 / mm
动态可学习参数大小	Avg	骨骼	骨折面	Avg	骨骼	骨折面
1×1×1	65.03	68.09	61.96	4.34	4.23	4.45
2×2×2	68.55	70.30	66.79	2.04	2.27	1.81
3×3×3	66.28	68.76	63.80	3.10	2.81	3.38
4×4×4	65.58	68.30	62.85	3.72	2.66	4.77
5×5×5	66.57	70.09	63.04	3.54	2.31	4.76

下载: 导出CSV

表 5 不同小波基对分割精度的影响

小波基类型	Dice Score/%			HD95 / mm
小波基类型	Avg	骨骼	骨折面	Avg	骨骼	骨折面
Haar	67.14	67.98	66.29	1.75	1.69	1.80
Daubechies	67.91	68.20	67.62	1.66	1.57	1.75
Coiflets	67.36	69.24	65.47	1.71	1.49	1.92
Biorthogonal	68.33	69.54	67.12	1.69	1.35	2.02
Symlets	67.50	67.21	67.79	1.84	1.92	1.75
Reverse Biorthogonal	69.35	71.64	67.06	1.58	1.27	1.89

下载: 导出CSV

表 6 GASG中不同轴解顺序对分割精度的影响

轴解顺序	Dice Score /%			HD95 / mm
轴解顺序	Avg	骨骼	骨折面	Avg	骨骼	骨折面
D→H→W	71.20	72.91	69.48	1.38	1.22	1.54
D→W→H	69.75	72.10	67.39	1.73	1.69	1.76
H→D→W	69.08	70.89	67.28	1.54	1.36	1.71
H→W→D	68.24	70.12	66.35	1.70	1.60	1.80
W→D→H	67.11	67.28	66.93	3.34	4.90	1.77
W→H→D	69.64	71.13	68.14	1.54	1.36	1.72

下载: 导出CSV

表 7 不同损失权重比例对分割精度的影响

损失权重比例	Dice Score /%			HD95 / mm
损失权重比例	Avg	骨骼	骨折面	Avg	骨骼	骨折面
1:1:1	70.64	73.40	67.88	2.35	1.20	3.50
0.7:1:5	70.27	72.47	68.06	1.44	1.28	1.59
0.5:1:3	70.27	73.06	67.47	1.63	1.47	1.79
0.5:1:5	71.20	72.91	69.48	1.38	1.22	1.54
0.5:1:7	69.70	71.45	67.94	1.65	1.64	1.65
0.3:1:5	71.13	73.07	69.18	1.43	1.28	1.57

下载: 导出CSV

参考文献(28)

[1]	JUTO H, NILSSON H, and MORBERG P. Epidemiology of adult ankle fractures: 1756 cases identified in Norrbotten County during 2009–2013 and classified according to AO/OTA[J]. BMC Musculoskeletal Disorders, 2018, 19(1): 441. doi: 10.1186/s12891-018-2326-x.
[2]	HARDALAÇ F, UYSAL F, PEKER O, et al. Fracture detection in wrist X-ray images using deep learning-based object detection models[J]. Sensors, 2022, 22(3): 1285. doi: 10.3390/s22031285.
[3]	BARG A, PAGENSTERT G I, HÜGLE T, et al. Ankle osteoarthritis: Etiology, diagnostics, and classification[J]. Foot and Ankle Clinics, 2013, 18(3): 411–426. doi: 10.1016/j.fcl.2013.06.001.
[4]	LI Mengnai, COLLIER R C, HILL B W, et al. Comparing different surgical techniques for addressing the posterior malleolus in supination external rotation ankle fractures and the need for syndesmotic screw fixation[J]. The Journal of Foot and Ankle Surgery, 2017, 56(4): 730–734. doi: 10.1053/j.jfas.2017.01.053.
[5]	ZHANG Qida, CHEN Zhenxian, PENG Yinghu, et al. The novel magnesium–titanium hybrid cannulated screws for the treatment of vertical femoral neck fractures: Biomechanical evaluation[J]. Journal of Orthopaedic Translation, 2023, 42: 127–136. doi: 10.1016/j.jot.2023.08.003.
[6]	SZYMAŃSKI T and ZDANOWICZ U. Comparison of routine computed tomography and plain X-ray imaging for malleolar fractures—How much do we miss?[J]. Foot and Ankle Surgery, 2022, 28(2): 263–268. doi: 10.1016/j.fas.2021.03.025.
[7]	ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S S, et al. 3D U-Net: Learning dense volumetric segmentation from sparse annotation[C]. Proceedings of the 19th International Conference on Medical Image Computing and Computer-Assisted Intervention, Athens, Greece, 2016: 424–432. doi: 10.1007/978-3-319-46723-8_49.
[8]	白静, 杨瞻源, 彭斌, 等. 三维卷积神经网络及其在视频理解领域中的应用研究[J]. 电子与信息学报, 2023, 45(6): 2273–2283. doi: 10.11999/JEIT220596. BAI Jing, YANG Zhanyuan, PENG Bin, et al. Research on 3D convolutional neural network and its application on video understanding[J]. Journal of Electronics & Information Technology, 2023, 45(6): 2273–2283. doi: 10.11999/JEIT220596.
[9]	LUO Xiaoling, WANG Wei, XU Yong, et al. A deep convolutional neural network for diabetic retinopathy detection via mining local and long‐range dependence[J]. CAAI Transactions on Intelligence Technology, 2024, 9(1): 153–166. doi: 10.1049/cit2.12155.
[10]	YAN Qingsen, LIU Shengqiang, XU Songhua, et al. 3D medical image segmentation using parallel transformers[J]. Pattern Recognition, 2023, 138: 109432. doi: 10.1016/j.patcog.2023.109432.
[11]	HUANG Yonghao, CHEN Leiting, ZHOU Chuan, et al. Model long-range dependencies for multi-modality and multi-view retinopathy diagnosis through transformers[J]. Knowledge-Based Systems, 2023, 271: 110544. doi: 10.1016/j.knosys.2023.110544.
[12]	CHEN Shuchao, LUO Cao, LIU Shanshan, et al. LD-UNet: A long-distance perceptual model for segmentation of blurred boundaries in medical images[J]. Computers in Biology and Medicine, 2024, 171: 108120. doi: 10.1016/j.compbiomed.2024.108120.
[13]	ZHANG Zhenguang, PENG Bo, and ZHAO Tingyu. An ultra-lightweight network combining Mamba and frequency-domain feature extraction for pavement tiny-crack segmentation[J]. Expert Systems with Applications, 2025, 264: 125941. doi: 10.1016/j.eswa.2024.125941.
[14]	HE Zhili, CHEN Wang, ZHANG Jian, et al. Crack segmentation on steel structures using boundary guidance model[J]. Automation in Construction, 2024, 162: 105354. doi: 10.1016/j.autcon.2024.105354.
[15]	CHU Honghu, WANG Wei, and DENG Lu. Tiny‐Crack‐Net: A multiscale feature fusion network with attention mechanisms for segmentation of tiny cracks[J]. Computer‐Aided Civil and Infrastructure Engineering, 2022, 37(14): 1914–1931. doi: 10.1111/mice.12881.
[16]	HUANG Yaqi, HU Ge, JI Changjin, et al. Glass-cutting medical images via a mechanical image segmentation method based on crack propagation[J]. Nature Communications, 2020, 11(1): 5669. doi: 10.1038/s41467-020-19392-7.
[17]	SUN Rui, LI Xuming, ZHANG Libing, et al. Wavelet-integrated deep neural network for deblurring and segmentation of crack images[J]. Mechanical Systems and Signal Processing, 2025, 224: 112240. doi: 10.1016/j.ymssp.2024.112240.
[18]	李文轩, 冯斌, 吴福理, 等. 基于纹理和几何特征融合的髁突骨微结构分割与预测[J]. 计算机辅助设计与图形学学报, 2025, 37(2): 340–348. doi: 10.3724/SP.J.1089.2023-00173. LI Wenxuan, FENG Bin, WU Fuli, et al. Bone microstructure segmentation and prediction of condyle based on fusion of texture and geometric features[J]. Journal of Computer-Aided Design & Computer Graphics, 2025, 37(2): 340–348. doi: 10.3724/SP.J.1089.2023-00173.
[19]	WANG Kun, ZHANG Xiaohong, LU Yuting, et al. GSAL: Geometric structure adversarial learning for robust medical image segmentation[J]. Pattern Recognition, 2023, 140: 109596. doi: 10.1016/j.patcog.2023.109596.
[20]	ISENSEE F, JAEGER P F, KOHL S A A, et al. nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation[J]. Nature Methods, 2021, 18(2): 203–211. doi: 10.1038/s41592-020-01008-z.
[21]	XIAO Ruoxiu, QI Siyu, REN Huayang, et al. Multi-objective constraints for path planning in screw fixation of scaphoid fractures[J]. Computers in Biology and Medicine, 2024, 182: 109163. doi: 10.1016/j.compbiomed.2024.109163.
[22]	LIU Jiaxuan, LI Haitao, ZENG Bolun, et al. An end-to-end geometry-based pipeline for automatic preoperative surgical planning of pelvic fracture reduction and fixation[J]. IEEE Transactions on Medical Imaging, 2025, 44(1): 79–91. doi: 10.1109/TMI.2024.3429403.
[23]	GU Hanxue, COLGLAZIER R, Dong Haoyu, et al. SegmentAnyBone: A universal model that segments any bone at any location on MRI[J]. Medical Image Analysis, 2025, 101: 103469. doi: 10.1016/j.media.2025.103469.
[24]	LU Xu, CUI Zengzhen, SUN Yihua, et al. Better rough than scarce: Proximal femur fracture segmentation with rough annotations[J]. IEEE Transactions on Medical Imaging, 2024, 43(9): 3240–3252. doi: 10.1109/TMI.2024.3392854.
[25]	ZHOU Hongyu, GUO Jiansen, ZHANG Yinghao, et al. nnFormer: Volumetric medical image segmentation via a 3D transformer[J]. IEEE Transactions on Image Processing, 2023, 32: 4036–4045. doi: 10.1109/TIP.2023.3293771.
[26]	SHAKER A, MAAZ M, RASHEED H, et al. UNETR++: Delving into efficient and accurate 3D medical image segmentation[J]. IEEE Transactions on Medical Imaging, 2024, 43(9): 3377–3390. doi: 10.1109/TMI.2024.3398728.
[27]	MA Jun, LI Feifei, and WANG Bo. U-mamba: Enhancing long-range dependency for biomedical image segmentation[J]. arXiv preprint arXiv: 2401.04722, 2024. doi: 10.48550/arXiv.2401.04722. (查阅网上资料,请核对文献类型及格式).
[28]	ZHOU Yanfeng, LI Lingrui, LU Le, et al. nnWNet: Rethinking the use of transformers in biomedical image segmentation and calling for a unified evaluation benchmark[C]. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA, 2025: 20852–20862. doi: 10.1109/CVPR52734.2025.01942.