Low-Dose Computed Tomography Image Denoising Method Based on Cross Domain and High Receptive Field

Denoising network	Dilation factor	Receptive field	Kernel size
Total receptive field		59
Dilated convolution+ReLU 1	1	3	3×3×128
Dilated convolution+ReLU 2	2	5	3×3×128
Dilated convolution+ReLU 3	3	7	3×3×128
Dilated convolution+ReLU 4	4	9	3×3×128
Dilated convolution+ReLU 5	5	11	3×3×128
Dilated convolution+ReLU 6	4	9	3×3×128
Dilated convolution+ReLU 7	3	7	3×3×128
Dilated convolution+ReLU 8	2	5	3×3×128
Dilated convolution+ReLU 9	1	3	3×3×128

Table 2. The number of CT slices from 10 patients in the AAPM dataset
View table
Table 2. The number of CT slices from 10 patients in the AAPM dataset
Patient ID 1 mm slice 3 mm slice
Total 5936 2378
L067 560 224
L096 823 330
L109 318 128
L143 585 234
L192 600 240
L286 525 210
L291 856 343
L310 533 214
L333 610 244
L506 526 211

Table 3. Comparison of LDTNet with different structures

View table

Table 3. Comparison of LDTNet with different structures

Structure No.	DTCWT	Network structure	Receptive field	Model size /MB	PSNR /dB
1	One-level	J0-Net+J1-Net	59×2	9.6	28.7504
2	One-level	J0-Net+J1-Net+X	59×2	9.6	28.7538
3	One-level	J0-Net+J1-Net+X-Net	59×2	14.4	28.7805
4	One-level	J0-Net+J1-Net+X-Net+X	59×2	14.4	28.7820
5	Two-level	J0-Net+J1-Net+J2-Net	59×3	14.4	28.6849
6	Two-level	J0-Net+J1-Net+J2-Net+X	59×3	14.4	28.6889
7	Two-level	J0-Net+J1-Net+J2-Net+X-Net	59×3	19.2	28.7818
8	Two-level	J0-Net+J1-Net+J2-Net+X-Net+X	59×3	19.2	28.7862
9	Three-level	J0-Net+J1-Net+J2-Net+J3-Net	59×4	19.2	28.6113
10	Three-level	J0-Net+J1-Net+J2-Net+J3-Net+X	59×4	19.2	28.6166
11	Three-level	J0-Net+J1-Net+J2-Net+J3-Net+X-Net	59×4	25.0	28.7839
12	Three-level	J0-Net+J1-Net+J2-Net+J3-Net+X-Net+X	59×4	25.0	28.7892

Table 4. The comparison of average PSNR for different filter combinations

View table

Table 4. The comparison of average PSNR for different filter combinations

Filter	Near Symmetric A	Near Symmetric B
Qshift 06	28.7375	28.7770
Qshift A	28.7638	28.7785
Qshift B	28.7670	28.7892
Qshift C	28.7867	28.7685
Qshift D	28.7787	28.7755
Filter	Antonini	LeGall
Qshift 06	28.7683	28.7712
Qshift A	28.7606	28.7785
Qshift B	28.7816	28.7601
Qshift C	28.7809	28.7723
Qshift D	28.7752	28.7880

Table 5. The comparison of average PSNR and SSIM for different methods on AAPM-1 mm dataset

View table

Table 5. The comparison of average PSNR and SSIM for different methods on AAPM-1 mm dataset

Method	PSNR /dB	SSIM
LDCT	18.8504	0.6125
CNN10	24.6883 ± 0.5066	0.6655 ± 0.0128
AAPM-Net	24.7552 ± 0.5543	0.6719 ± 0.0133
RED-CNN	24.7671 ± 0.5681	0.6723 ± 0.0151
DRN-LDCT	24.8241 ± 0.5203	0.6733 ± 0.0149
MDTNet	24.8650 ± 0.5339	0.6749 ± 0.0130
LDTNet	24.8859 ± 0.5102	0.6760 ± 0.0135

Table 6. The comparison of average PSNR and SSIM for different methods on AAPM-3 mm dataset

View table

Table 6. The comparison of average PSNR and SSIM for different methods on AAPM-3 mm dataset

Method	PSNR /dB	SSIM
LDCT	22.9338	0.6732
CNN10	28.6361 ± 0.5510	0.7478 ± 0.0012
AAPM-Net	28.6553 ± 0.5777	0.7480 ± 0.0099
RED-CNN	28.7254 ± 0.5811	0.7512 ± 0.0153
DRN-LDCT	28.7284 ± 0.5602	0.7496 ± 0.0125
MDTNet	28.7555 ± 0.5674	0.7517 ± 0.0110
LDTNet	28.7892 ± 0.5545	0.7633 ± 0.0095

Table 7. Comparison of different methods
View table
Table 7. Comparison of different methods
Method Model size /MB Total time /s Average time /s
MDTNet 16.0 3.85 0.193
DRN-LDCT 23.9 5.01 0.251
LDTNet 25.0 5.25 0.263

Tools

Get Citation

Copy Citation Text

Lu Zhang, Pengju Liu, Shigang Liu. Low-Dose Computed Tomography Image Denoising Method Based on Cross Domain and High Receptive Field[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1417003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites