Large Kernel Convolution and Transformer Parallelism Based 3D Medical Image Registration Modeling

Fig. 7. Registration results on the OASIS dataset. (a) The fixed image to be registered; (b) the moving image to be registered; (c) registration of ANTs; (d) registration of VoxelMorph-V1; (e) registration of VoxelMorph-V2; (f) registration of SYM-Net; (g) registration of SymTrans; (h) registration of TransMorph; (i) registration of TransMatch; (j) registration of our algorithm

Download full size

Fig. 8. Comparison of ablation experiment. (a) The fixed image to be registered; (b) the moving image to be registered; (c) registration of VoxelMorph-V1; (d) registration of LKC Block; (e) registration of 3D Swin Transformer; (f) registration of MSAA; (g) registration of PLKCT-UNet

Download full size

Fig. 9. Comparison of generalization. (a) The fixed image to be registered; (b) the moving image to be registered; (c) registration of ANTs; (d) registration of VoxelMorph-V1; (e) registration of VoxelMorph-V2; (f) registration of SYM-Net; (g) registration of SymTrans; (h) registration of TransMorph; (i) registration of TransMatch; (j) registration of PLKCT-UNet

Download full size

Table 1. Environment configuration
View table
Table 1. Environment configuration
Environment Version
GPU RTX3090（24 GB）
CUDA 11.8
Python 3.8
PyTorch 2.0.1

Table 2. Parameters setting
View table
Table 2. Parameters setting
Parameter setting Value
Epoch 500
Iterations 145000
Batch_size 1
Learning rate 1×10^-4
λ 4.0
γ 0.67
Optimizer Adam

Table 3. Model comparison results

View table

Table 3. Model comparison results

Model	Parameters /10⁶	Dice	HD	Time /s
ANTs（SyN）		0.783	3.41	33.68
VoxelMorph-V1	0.36	0.782	2.68	0.26
VoxelMorph-V2	0.39	0.794	2.62	0.28
SYM-Net	4.47	0.789	2.64	0.36
SymTrans	16.05	0.805	2.56	0.32
TransMorph	107.94	0.818	2.49	0.43
TransMatch	112.41	0.815	2.53	0.39
Ours	86.64	0.824	2.46	0.34

Table 4. Results of the ablation experiment
View table
Table 4. Results of the ablation experiment
Model LKC Swin Transformer MSAA Dice HD
VoxelMorph × × × 0.794 2.62
√ × × 0.805 2.55
× √ × 0.812 2.53
× × √ 0.799 2.58
√ √ √ 0.824 2.47

Table 5. Comparison of generalization

View table

Table 5. Comparison of generalization

Model	Dice	HD	Time /s
ANTs （SyN）	0.670	6.724	27.84
VoxelMorph-v1	0.679	6.698	0.25
VoxelMorph-v2	0.685	6.581	0.34
SYM-Net	0.697	6.475	0.31
SymTrans	0.703	6.432	0.39
TransMorph	0.712	6.369	0.37
TransMatch	0.709	6.383	0.34
PLKCT-UNet	0.716	6.357	0.32

Tools

Get Citation

Copy Citation Text

Jing Peng, Jiarong Yan, Yu Shen, Jiaying Liu, Ziyi Wei, Shan Bai, Jiangcheng Li, Yukun Ma, Ruoxuan Wang. Large Kernel Convolution and Transformer Parallelism Based 3D Medical Image Registration Modeling[J]. Chinese Journal of Lasers, 2025, 52(3): 0307104

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites