2D-3D Medical Image Registration Based on Training-Inference Decoupling Architecture

Table 1. Transformation parameter distribution
View table
Table 1. Transformation parameter distribution
Parameter Distribution range
t_x /mm U（-20， 20）
t_y /mm U（-20， 20）
t_z /mm U（-10， 10）
t_θ /（°） U（-10， 10）
t_α /（°） U（-10， 10）
t_β /（°） U（-5， 5）

Table 2. Number distribution of the datasets
View table
Table 2. Number distribution of the datasets
Structure Number of images /10⁴
Training set Test set Data set
Pelvis 3 1 4
Chest 3 1 4

Table 3. Comparison of network hyperparameters
View table
Table 3. Comparison of network hyperparameters
Batchsize T-MAE /mm R-MAE /（°） RMSE
8 0.08 0.05 0.08
16 0.10 0.06 0.09
32 0.15 0.12 0.12

Table 4. Error comparison of branch structure ablation experiments

View table

Table 4. Error comparison of branch structure ablation experiments

Dataset	Network structure	T-MAE /mm	R-MAE /（°）	RMSE
Pelvis	No branch	0.103	0.073	0.115
	Identity	0.087	0.062	0.098
	1×1 Conv	0.083	0.057	0.093
	Proposed method	0.075	0.051	0.084
Chest	No branch	0.155	0.099	0.169
	Identity	0.147	0.088	0.156
	1×1 Conv	0.129	0.082	0.143
	Proposed method	0.115	0.068	0.125

Table 5. Time comparison of branch structure ablation experiments
View table
Table 5. Time comparison of branch structure ablation experiments
Network structure Parameter /10⁶ T-time /ms I-time /ms
No branch 50.98 29.8 25.7
Identity 50.98 30.2 25.8
1×1 Conv 56.53 33.2 25.8
Proposed method 56.53 34.6 26.0

Table 6. Activation function comparison

View table

Table 6. Activation function comparison

Dataset	Activation function	T-MAE /mm	R-MAE /（°）	RMSE
Pelvis	ReLU	0.078	0.052	0.087
	Swish	0.079	0.053	0.088
	ACON-C	0.078	0.052	0.087
	Meta-ACON	0.075	0.051	0.084
Chest	ReLU	0.143	0.084	0.156
	Swish	0.151	0.082	0.162
	ACON-C	0.129	0.077	0.129
	Meta-ACON	0.115	0.068	0.125

Table 7. Comparison of registration effects of different networks

View table

Table 7. Comparison of registration effects of different networks

Dataset	Network structure	T-MAE	R-MAE	RMSE	Time /ms
Pelvis	GoogLeNet	0.36±0.19	0.23±0.13	0.39±0.24	16.0
	VGG16	0.32±0.27	0.21±0.18	0.35±0.21	19.2
	ResNet50	0.09±0.08	0.06±0.07	0.10±0.08	18.4
	DenseNet121	0.11±0.08	0.07±0.06	0.12±0.09	34.5
	Net+GradNCC^［11］	7.83±19.8	4.94±8.78
	Proposed method	0.08±0.04	0.05±0.03	0.08±0.06	26.0
Chest	GoogLeNet	0.41±0.18	0.16±0.09	0.39±0.20	21.5
	VGG16	0.56±0.45	0.31±0.24	0.60±0.41	19.1
	ResNet50	0.33±0.09	0.20±0.05	0.35±0.12	24.6
	DenseNet121	0.15±0.08	0.09±0.05	0.16±0.08	37.0
	Inception+branch^［10］	0.61	1.93		2.3
	Proposed method	0.12±0.07	0.07±0.04	0.13±0.08	26.4

Table 8. Comparison of registration methods

View table

Table 8. Comparison of registration methods

Dataset	Method	NCC	NMI	SSIM	Time /ms
Pelvis	NCC_Powell	0.56±0.07	0.16±0.02	0.57±0.09	96.4×10³
	NMI_Powell	0.61±0.09	0.15±0.01	0.57±0.09	28.6×10³
	Deep learning method^［9］	0.82±0.07	0.32±0.03		30.0
	Proposed method	0.99±0.05	0.72±0.03	0.99±0.08	26.0
Chest	NCC_Powell	0.53±0.11	0.11±0.02	0.55±0.09	101.5×10³
	NMI_Powell	0.57±0.09	0.11±0.02	0.58±0.08	29.2×10³
	Inception+branch^［10］			0.42	2.3×10³
	Proposed method	0.99±0.06	0.56±0.02	0.99±0.09	26.4

Tools

Get Citation

Copy Citation Text

Wenjü Li, Deqing Kong, Guogang Cao, Sicheng Li, Cuixia Dai. 2D-3D Medical Image Registration Based on Training-Inference Decoupling Architecture[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1610015

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites