Scattering Correction of CBCT Images Based on Dual Encoding U-Net and Discrete Wavelet Transform

Ting Yan; Chaoliang He; Qian Tong; Songtao Zhu; Xiaojiao Duan

doi:10.3788/AOS241966

Acta Optica Sinica, Volume. 45, Issue 9, 0910001(2025)

Scattering Correction of CBCT Images Based on Dual Encoding U-Net and Discrete Wavelet Transform

Ting Yan^1,2, Chaoliang He^1,2, Qian Tong^1,2, Songtao Zhu^1,2, and Xiaojiao Duan^1,2、*

¹Key Laboratory of Optoelectronic Technology & Systems, Ministry of Education, Chongqing University, Chongqing 400044, China

²Industrial CT Non-Destructive Testing Engineering Research Center, Ministry of Education, Chongqing University, Chongqing 400044, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(19)

Fig. 1. DEU-Net structure

Download full size

Fig. 2. Network modules. (a) DCCNN block; (b) Swin-T block

Download full size

Fig. 3. Scattering correction process

Download full size

Fig. 4. Curve of loss function variation

Download full size

Fig. 5. Comparison of results before and after correction. (a1)(b1) Original images; (a2)(b2) label images; (a3)(b3) corrected images; (a4)(b4) CT values at the same position of three data; (a5)(b5) difference plots of CT values obtained from Figs. 5(a1) and 5(a2) and from Figs. 5(b1) and 5(b2); (a6)(b6) difference plots of CT values obtained from Figs. 5(a2) and 5(a3) and from Figs. 5(b2) and 5(b3)

Download full size

Fig. 6. Comparison with other methods. (a1) Uncorrected image; (b1) label image; (c1) image obtained by U-Net correction; (d1) image obtained by DRCNN correction; (e1) image obtained by P2PGAN correction; (f1) image obtained by CycleGAN correction; (g1) image obtained by combining DEU-Net with scattering correction block; (a2)‒(g2) difference maps of CT values obtained by subtracting label data from data obtained by correction using different methods

Download full size

Fig. 7. Ablation experimental results. (a1) Uncorrected image; (b1) labeled image; (c1) DCCNN corrected image; (d1) corrected image of DCCNN combined with Swin-T; (e1) corrected image of DCCNN combined with Swin-T and weighted loss function; (f1) corrected image of DEU-Net model combined with scattering correction module; (a2)‒(f2) differences between CT values obtained by scattering correction by different methods and label data

Download full size

Fig. 8. Correction results of different post-processing methods. (a1) Uncorrected image; (b1) labeled image; (c1) corrected image of 3×3 low-pass filtering; (d1) corrected image of 5×5 low-pass filtering; (e1) corrected image of 9×9 low-pass filtering; (f1) corrected image of DWT; (a2)‒(f2) differences between CT values obtained using different post-processing methods for scattering correction and labeled data

Download full size

Fig. 9. Correction results of different post-processing methods

Download full size

Fig. 10. Comparison of low-dose CT images before and after correction. (a1)(b1)(c1) Uncorrected images; (a2)(b2)(c2) label images; (a3)(b3)(c3) corrected images; (a4)(b4)(c4) CT value comparisons of three images on the same horizontal line

Download full size

Fig. 11. Comparison of CT values of ROIs of fat and muscle before and after correction. (a1)(b1) Uncorrected images; (a2)(b2) labeled images; (a3)(b3) corrected images; (a4)(b4) comparison of muscle CT values of four ROIs in three images; (a5)(b5) comparison of fat CT values of four ROIs in three images

Download full size

Fig. 12. CBCT system

Download full size

Fig. 13. Comparison of turbine blade correction effects. (a1)(b1) Uncorrected images; (a2)(b2) labeled images; (a3)(b3) images corrected by BSG; (a4)(b4) images corrected using method in this paper; (a5)(b5) comparison of gray values of the same horizontal line in four images

Download full size

Table 1. Indicators of different correction methods
View table
Table 1. Indicators of different correction methods
Method PSNR SSIM RMSE /HU
Uncorrected 19.8550 0.6875 102.6088
U-Net 36.5163 0.9616 15.0779
DRCNN 38.9564 0.9749 11.3299
P2PGAN 32.3787 0.9550 24.0471
CycleGAN 33.6433 0.9665 20.7891
Ours 42.8761 0.9857 7.2167

Table 2. Ablation experimental data indicators
View table
Table 2. Ablation experimental data indicators
Methods DCCNN Swin-T W-Loss DWT PSNR SSIM RMSE /HU
M1 √ 39.3441 0.9728 10.9523
M2 √ √ 40.2947 0.9788 9.7143
M3 √ √ √ 40.8323 0.9808 9.1376
M4 （ours） √ √ √ √ 42.8761 0.9857 7.2167

Table 3. Comparison of data before and after correction
View table
Table 3. Comparison of data before and after correction
Method PSNR SSIM RMSE /HU
Uncorrected 14.1879 0.5401 195.3284
Ours 36.0679 0.9056 16.0692
Reference — 1.0000 0

Table 4. CT data corresponding to Figs. 11(a1)‒(a3)

View table

Table 4. CT data corresponding to Figs. 11(a1)‒(a3)

ROI	Evaluation	Uncorrected	Corrected	Ground truth
	Mean	-150.7834	45.5839	61.0119
Muscle	Standard	30.8970	11.4375	18.9890
	RMSE	213.5861	23.2215	0
	Mean	-311.0819	-114.9116	-100.5558
Fat	Standard	26.8770	14.6077	18.6641
	RMSE	211.7512	19.4458	0

Table 5. Contrast data of ROIs before and after correction
View table
Table 5. Contrast data of ROIs before and after correction
ROI Uncorrected BSG Ours Label
ROI 1 139.3850 158.7775 207.6525 206.4689
ROI 2 113.7350 129.9975 195.3225 194.7114
ROI 3 128.8683 135.3600 206.6944 207.6606
ROI 4 96.5450 111.1600 191. 0950 191. 8000

Table 6. CNR data of ROIs before and after correction
View table
Table 6. CNR data of ROIs before and after correction
ROI Uncorrected BSG Ours Label
ROI 1 28.5362 34.1308 95.5111 67.6825
ROI 2 17.6328 25.3179 63.9055 50.8019
ROI 3 68.4396 80.2468 110.5621 114.0659
ROI 4 6.3913 14.7569 29. 5211 29. 6307

Tools

Get Citation

Copy Citation Text

Ting Yan, Chaoliang He, Qian Tong, Songtao Zhu, Xiaojiao Duan. Scattering Correction of CBCT Images Based on Dual Encoding U-Net and Discrete Wavelet Transform[J]. Acta Optica Sinica, 2025, 45(9): 0910001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites