基于多尺度膨胀残差和双重注意力的可变形配准网络

doi:10.11938/cjmr20253190

摘要/Abstract

摘要：

可变形配准在多项医学影像分析任务中具有重要作用．然而，卷积神经网络的固定感受野难以充分捕捉大脑的空间上下文信息，Transformer架构能有效捕获全局信息但计算开销大．因此，本文提出一种基于多尺度膨胀残差卷积和双重注意力的可变形配准网络．该网络采用多尺度膨胀残差卷积同时捕获局部细节和广泛的上下文信息，通过双重注意力模块增强特征的表达能力，并在解码器部分引入动态上采样精细重建高频细节以确保变形场的拓扑完整性．在两个公共脑部数据集上的实验结果表明，本文所提模型在戴斯相似性系数（DSC）、95%豪斯多夫距离（HD95）和非正雅可比行列式百分比等多个指标上实现了较高的配准效果，表明其具有更强的配准精度和更平滑的变形场．

关键词: 可变形图像配准, 多尺度, 双重注意力, 动态上采样

Abstract:

Deformable registration plays a significant role in multiple medical image analysis tasks. However, the fixed receptive field of convolutional neural networks makes it difficult to fully capture the spatial context information of the brain. While the Transformer architecture can effectively capture global information, it suffers from high computational cost. To address this dilemma, we propose a deformable registration network based on multi-scale dilated residual convolution and dual attention. This network employs a multi-scale dilated residual convolution to capture local details and extensive context information simultaneously. A dual attention module is integrated to enhance feature representation capability. The decoder section introduces dynamic upsampling to precisely reconstruct high-frequency details, thereby ensuring the topological integrity of the deformation field. Experiments were conducted on two public brain datasets. The results demonstrate that the proposed model achieves high registration performance across multiple metrics, including Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and percentage of non-positive Jacobian determinant, indicating superior registration precision and smoother deformation fields.

Key words: deformable image registration, multi-scale, dual attention, dynamic upsampling

中图分类号:

O482.53

羊晶晶, 王远军. 基于多尺度膨胀残差和双重注意力的可变形配准网络[J]. 波谱学杂志, 2026, 43(3): 291-306.

YANG Jingjing, WANG Yuanjun. Deformable Registration Network Based on Multi-scale Dilated Residual and Dual Attention[J]. Chinese Journal of Magnetic Resonance, 2026, 43(3): 291-306.

图/表 15

图1

图2

图3

图4

表1

表2

图5

图6

表3

图7

表4

表5

表A1

表A2

表A3

参考文献 35

[1]	ZHU H Y, LI M, JI Y L, et al. A body examination study of brain MRI registration in adolescents using fusion generative adversarial network[J]. Chinese J Magn Reson Imaging, 2023, 14(2): 116-124.
	朱海艳, 李盟, 季跃龙, 等. 融合生成对抗网络的青少年脑MRI配准体测研究[J]. 磁共振成像, 2023, 14(2): 116-124.
[2]	ZHOU Q, WANG Y J. Groupwise registration for magnetic resonance image based on variational inference[J]. Chinese J Magn Reson, 2022, 39(3): 291-302.
	周勤, 王远军. 基于变分推断的磁共振图像群组配准[J]. 波谱学杂志, 2022, 39(3): 291-302. doi: 10.11938/cjmr20212918
[3]	BAN Y, WANG Y, LIU S, et al. 2D/3D multimode medical image alignment based on spatial histograms[J]. Appl Sci, 2022, 12(16): 8261. doi: 10.3390/app12168261
[4]	BOSMA L S, ZACHIU C, SENNEVILLE B D D, et al. Intensity-based quality assurance criteria for deformable image registration in image-guided radiotherapy[J]. Med Phys, 2023, 50(9): 5715-5722. doi: 10.1002/mp.v50.9
[5]	XUE P, DONG E, JI H. Lung 4D CT image registration based on high-order markov random field[J]. IEEE Trans Med Imaging, 2019, 39(4): 910-921. doi: 10.1109/TMI.42
[6]	MANERA A L, DADAR M, FONOV V, et al. CerebrA, registration and manual label correction of Mindboggle-101 atlas for MNI-ICBM152 template[J]. Sci Data, 2020, 7(1): 237. doi: 10.1038/s41597-020-0557-9 pmid: 32669554
[7]	BALAKRISHNAN G, ZHAO A, SABUNCU M R, et al. Voxelmorph: a learning framework for deformable medical image registration[J]. IEEE Trans Med Imaging, 2019, 38(8): 1788-1800. doi: 10.1109/TMI.42
[8]	CHEN X, WANG X, ZHANG K, et al. Recent advances and clinical applications of deep learning in medical image analysis[J]. Med Image Anal, 2022, 79: 102444. doi: 10.1016/j.media.2022.102444
[9]	JADERBERG M, SIMONYAN K, ZISSERMAN A. Spatial transformer networks[J]. Adv Neural Inf Process Syst, 2015, 28: 2017-2025.
[10]	KIM B, KIM D H, PARK S H, et al. CycleMorph: cycle consistent unsupervised deformable image registration[J]. Med Image Anal, 2021, 71: 102036. doi: 10.1016/j.media.2021.102036
[11]	MOK T C W, CHUNG A. Fast symmetric diffeomorphic image registration with convolutional neural networks[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seattle, WA, USA: IEEE, 2020: 4644-4653.
[12]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]// Advances in Neural Information Processing Systems 30. Long Beach, CA, USA: Curran Associates, 2017: 5998-6008.
[13]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16×16 words:Transformers for image recognition at scale[PP/OL]. arXiv (2021-06-03)[2025-11-25]. https://arxiv.org/abs/2010.11929.
[14]	CHEN J, HE Y, FREY E C, et al. Vit-V-Net: Vision transformer for unsupervised volumetric medical image registration[PP/OL]. arXiv 2021-04-13)[2025-11-25]. https://arxiv.org/abs/2104.06468.
[15]	CHEN J, FREY E C, HE Y, et al. Transmorph: Transformer for unsupervised medical image registration[J]. Med Image Anal, 2022, 82: 102615. doi: 10.1016/j.media.2022.102615
[16]	LIU Z, LIN Y, CAO Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, QC, Canada: IEEE, 2021: 10012-10022.
[17]	SHI J, HE Y, KONG Y, et al. Xmorpher: Full transformer for deformable medical image registration via cross attention[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention. Singapore: Springer, 2022: 217-226.
[18]	ZHENG J Q, WANG Z, HUANG B, et al. Recursive deformable image registration network with mutual attention[C]// Annual Conference on Medical Image Understanding and Analysis. Cambridge, UK: Springer, 2022: 75-86.
[19]	WANG H, NI D, WANG Y. Modet: Learning deformable image registration via motion decomposition transformer[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention. Vancouver, BC, Canada: Springer, 2023: 740-749.
[20]	CHEN Z, ZHENG Y, GEE J C. Transmatch: A transformer-based multilevel dual-stream feature matching network for unsupervised deformable image registration[J]. IEEE Trans Med Imaging, 2023, 43(1): 15-27. doi: 10.1109/TMI.2023.3288136
[21]	LIU Z, MAO H, WU C Y, et al. A convnet for the 2020s[C]// Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. New Orleans, LA, USA: IEEE, 2022: 11976-11986.
[22]	CHOLLET F. Xception: Deep learning with depthwise separable convolutions[C]// Proceedings of the IEEE conference on computer vision and pattern recognition. Honolulu, HI, USA: IEEE, 2017: 1251-1258.
[23]	LIU W, LU H, FU H, et al. Learning to upsample by learning to sample[C]// Proceedings of the IEEE/CVF international conference on computer vision. Paris, France: IEEE, 2023: 6027-6037.
[24]	WANG J, CHEN K, XU R, et al. Carafe: Content-aware reassembly of features[C]// Proceedings of the IEEE/CVF international conference on computer vision. Seoul, Korea: IEEE, 2019: 3007-3016.
[25]	LU H, LIU W, FU H, et al. FADE: Fusing the assets of decoder and encoder for task-agnostic upsampling[C]// European conference on computer vision. Tel Aviv, Israel: Springer, 2022: 231-247.
[26]	RAO Y R, PRATHAPANI N, NAGABHOOSHANAM E. Application of normalized cross correlation to image registration[J]. Int J Res Eng Technol, 2014, 3(5): 12-16.
[27]	FISCHL B. FreeSurfer[J]. NeuroImage, 2012, 62(2): 774-781. doi: 10.1016/j.neuroimage.2012.01.021 pmid: 22248573
[28]	SHATTUCK D W, MIRZA M, ADISETIYO V, et al. Construction of a 3D probabilistic atlas of human cortical structures[J]. NeuroImage, 2008, 39(3): 1064-1080. doi: 10.1016/j.neuroimage.2007.09.031 pmid: 18037310
[29]	AVANTS B B, EPSTEIN C L, GROSSMAN M, et al. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain[J]. Med Image Anal, 2008, 12(1): 26-41. doi: 10.1016/j.media.2007.06.004 pmid: 17659998
[30]	MODAT M, RIDGWAY G R, TAYLOR Z A, et al. Fast free-form deformation using graphics processing units[J]. Comput Meth Prog Bio, 2010, 98(3): 278-284. doi: 10.1016/j.cmpb.2009.09.002 pmid: 19818524
[31]	DICE L R. Measures of the amount of ecologic association between species[J]. Ecology, 1945, 26(3): 297-302. doi: 10.2307/1932409
[32]	TAHA A A, HANBURY A. An efficient algorithm for calculating the exact Hausdorff distance[J]. IEEE Trans Pattern Anal Mach Intell, 2015, 37(11): 2153-2163. doi: 10.1109/TPAMI.2015.2408351
[33]	CHEN L C, ZHU Y, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]// Proceedings of the European conference on computer vision. Munich, Germany: Springer, 2018: 801-818.
[34]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional block attention module[C]// Proceedings of the European conference on computer vision. Munich, Germany: Springer, 2018: 3-19.
[35]	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]// Proceedings of the IEEE conference on computer vision and pattern recognition. Salt Lake City, UT, USA: IEEE, 2018: 7132-7141.

Model	DSC	HD95/mm	%\|J_ϕ\|≤0/%	Params/k	Flops/G	Memory/MB	Time/s
SyN	0.641 ± 0.040	2.118 ± 0.754	<0.001	N/A	N/A	254	189.172
NiftyReg	0.635 ± 0.066	2.166 ± 0.655	0.006 ± 0.016	N/A	N/A	254	69.962
VoxelMorph-1	0.722 ± 0.029	2.287 ± 0.805	1.668 ± 0.345	274.39	305.32	9128	0.378
VoxelMorph-2	0.728 ± 0.030	2.278 ± 0.797	1.580 ± 0.344	301.41	400.31	10524	0.417
CycleMorph	0.738 ± 0.028	2.156 ± 0.760	1.732 ± 0.365	361.30	127.02	23724	0.483
Vit-V-Net	0.643 ± 0.041	2.399 ± 0.854	1.391 ± 0.284	31560.08	391.12	11370	0.317
TransMorph	0.744 ± 0.031	2.273 ± 0.786	1.540 ± 0.335	46771.25	714.16	16840	0.450
MDDA-Net （本文模型）	0.751 ± 0.027	2.107 ± 0.765	1.384 ± 0.355	439.18	581.96	16572	0.364

Model	DSC	HD95/mm	%\|J_ϕ\|≤0/%	Params/k	Flops/G	Memory/MB	Time/s
SyN	0.670 ± 0.032	1.105 ± 0.180	<0.001	N/A	N/A	204	108.688
NiftyReg	0.660 ± 0.018	1.312 ± 0.401	<0.001	N/A	N/A	204	32.996
VoxelMorph-1	0.650 ± 0.027	1.310 ± 0.399	0.717 ± 0.302	274.39	218.09	8080	0.207
VoxelMorph-2	0.651 ± 0.029	1.143 ± 0.261	0.665 ± 0.285	301.41	285.94	9804	0.242
CycleMorph	0.670 ± 0.026	1.143 ± 0.261	0.625 ± 0.236	361.30	90.73	17192	0.221
Vit-V-Net	0.636 ± 0.029	1.258 ± 0.409	0.586 ± 0.268	31560.08	279.31	8044	0.225
TransMorph	0.668 ± 0.027	1.177 ± 0.339	0.630 ± 0.240	46771.25	511.97	11978	0.257
MDDA-Net （本文模型）	0.673 ± 0.023	1.104 ± 0.179	0.507 ± 0.139	439.18	415.68	11954	0.237

MDRC	CSDA	Dysample	DSC	HD95/mm	%\|J_ϕ\|≤0/%	Params/k	Flops/G	Memory/MB	Time/s
			0.730 ± 0.030	2.146 ± 0.809	1.645 ± 0.355	396.55	571.33	11052	0.213
√			0.737 ± 0.028	2.305 ± 0.843	1.567 ± 0.349	413.41	575.03	13230	0.212
	√		0.734 ± 0.029	2.197 ± 0.789	1.565 ± 0.348	397.36	572.62	12558	0.208
		√	0.745 ± 0.027	2.201 ± 0.786	1.533 ± 0.332	421.51	576.96	12910	0.211
√	√		0.739 ± 0.029	2.162 ± 0.793	1.496 ± 0.344	414.22	576.33	14732	0.216
√		√	0.749 ± 0.026	2.134 ± 0.771	1.383 ± 0.328	438.37	580.66	15068	0.339
	√	√	0.748 ± 0.028	2.131 ± 0.767	1.523 ± 0.343	422.32	578.25	14298	0.235
√	√	√	0.751 ± 0.027	2.107 ± 0.765	1.384 ± 0.355	439.18	581.96	16572	0.364

Model	DSC	HD95/mm	%\|J_ϕ\|≤0/%	Params/k	Flops/G	Memory/MB	Time/s(GPU)
+ASPP	0.751 ± 0.028	2.124 ± 0.782	1.506 ± 0.352	466.05	591.62	18978	0.374
+MDRC	0.751 ± 0.027	2.107 ± 0.765	1.384 ± 0.355	439.18	581.96	16572	0.364
+CBAM	0.745 ± 0.026	2.170 ± 0.826	1.448 ± 0.344	439.00	581.65	16310	0.377
+SE	0.747 ± 0.026	2.151 ± 0.834	1.394 ± 0.328	438.79	580.95	15630	0.344
+CSDA	0.751 ± 0.027	2.107 ± 0.765	1.384 ± 0.355	439.18	581.96	16572	0.364
+nearest	0.739 ± 0.029	2.162 ± 0.793	1.496 ± 0.344	414.22	576.33	14732	0.216
+trilinear	0.740 ± 0.023	2.126 ± 0.766	1.434 ± 0.379	414.22	576.33	14732	0.351
+Dysample	0.751 ± 0.027	2.107 ± 0.765	1.384 ± 0.355	439.18	581.96	16572	0.364

Dilation	DSC	HD95/mm	%\|J_ϕ\|≤0/%	Params/k	Flops/G	Memory/MB	Time/s (GPU)
1,2,1,2,2	0.753 ± 0.027	2.148 ± 0.820	1.541 ± 0.339	439.18	581.96	16572	0.363
1,1,2,2,2	0.749 ± 0.026	2.108 ± 0.737	1.393 ± 0.313	439.18	581.96	16572	0.364
1,1,1,2,3	0.750 ± 0.026	2.254 ± 0.789	1.398 ± 0.318	439.18	581.96	16572	0.364
1,1,1,2,2	0.751 ± 0.027	2.107 ± 0.765	1.384 ± 0.355	439.18	581.96	16572	0.364