Acta Optica Sinica, Volume. 44, Issue 5, 0515001(2024)
Dual-Energy CT Base Material Decomposition Method Based on Multi-Channel Cross-Convolution UCTransNet
Figures & Tables(20)
Fig. 3. Decomposition results of different methods in the Group 1 test set, where the first row represents images of bone-based material, and the second row represents images of soft tissue-based material
Fig. 4. Decomposition results of different methods in Group 2 test set, where the first row represents images of bone-based material, and the second row represents images of soft tissue-based material
Fig. 5. PSNR values of bone-based materials and soft tissue-based materials decomposition results of different methods in four test sets. (a) Bone-based materials; (b) soft tissue-based materials
Fig. 6. MAE, MSE, and SSIM values of bone-based materials and soft tissue-based materials decomposition results of different methods in four test sets. Among them, M1, M2, M3, M4, M5, and M6 represent Matrix Inversion, Iterative Decomposition, FCN, Butterfly-net, DIWGAN, and MC-UCTransNet, respectively. (a) Bone-based materials; (b) soft tissue-based materials
Fig. 7. Low-energy reconstruction images of decomposition results obtained by different methods. (a) Reference low-energy images; (b) reconstructed images of FCN; (c) reconstructed images of Butterfly-net; (d) reconstructed images of DIWGAN; (e) reconstructed images of MC-UCTransNet
Fig. 8. Difference between reference low energy images and decomposition results obtained by different methods. (a) FCN; (b) Butterfly-net; (c) DIWGAN; (d) MC-UCTransNet
Fig. 9. Decomposition results of different methods in the Group 1 test set, where the first row represents images of iodine-based material, and the second row represents images of soft tissue-based material
Fig. 10. Decomposition results of different methods in the Group 2 test set, where the first row represents images of iodine-based material, and the second row represents images of soft tissue-based material
Fig. 11. PSNR and SSIM in the test set for the loss function under different combinations of hyperparameters
Fig. 12. Convergence of loss and PSNR on different networks in training and validation sets. (a) Loss; (b) PSNR
Table 1. Quantitative scores of bone-based material decomposition results with different methods unit: mean±variance
View tableTable 1. Quantitative scores of bone-based material decomposition results with different methods unit: mean±variance
Method Matrix Inversion Iterative Decomposition FCN Butterfly-net DIWGAN MC-UCTransNet MAE 0.7676×10-2±4.13×10-8 0.7953×10-2±2.83×10-8 0.2218×10-2±2.40×10-8 0.1299×10-2±0.88×10-8 0.1022×10-2±0.65×10-8 0.0924×10-2±0.59×10-8 MSE 0.0479×10-2±2.26×10-9 0.0434×10-2±1.89×10-9 0.0366×10-2±1.66×10-9 0.0115×10-2±0.30×10-9 0.0108×10-2±0.22×10-9 0.0098×10-2±0.22×10-9 PSNR 32.6585±0.21 33.0787±0.20 33.8317±0.24 38.8924±0.47 39.1308±0.35 39.5849±0.46 SSIM 0.9701±4.02×10-6 0.9851±0.88×10-6 0.9807±0.94×10-6 0.9926±0.77×10-6 0.9934±0.54×10-6 0.9939±0.65×10-6
Table 2. Quantitative scores of soft tissue-based material decomposition results with different methods unit: mean±variance
View tableTable 2. Quantitative scores of soft tissue-based material decomposition results with different methods unit: mean±variance
Method Matrix Inversion Iterative Decomposition FCN Butterfly-net DIWGAN MC-UCTransNet MAE 2.7783×10-2±2.05×10-7 2.4921×10-2±2.30×10-7 0.5653×10-2±0.15×10-7 0.2742×10-2±0.06×10-7 0.2468×10-2±0.04×10-7 0.2650×10-2±0.04×10-7 MSE 0.2662×10-2±0.83×10-8 0.2719×10-2±1.03×10-8 0.0315×10-2±7.06×10-8 0.0116×10-2±0.008×10-8 0.0089×10-2±0.006×10-8 0.0062×10-2±0.003×10-8 PSNR 23.6462±0.02 23.5547±0.03 32.9346±0.14 37.2569±0.10 38.4406±0.13 39.9546±0.13 SSIM 0.8589±1.41×10-5 0.9321±0.18×10-5 0.9574±0.13×10-5 0.9871±0.03×10-5 0.9888±0.03×10-5 0.9904±0.02×10-5
Table 3. Quantitative scores of iodine-based material decomposition results of different methods unit: mean±variance
View tableTable 3. Quantitative scores of iodine-based material decomposition results of different methods unit: mean±variance
Method Matrix Inversion Iterative Decomposition FCN Butterfly-net DIWGAN MC-UCTransNet MAE 1.0684×10-2±73.29×10-8 1.0729×10-2±77.35×10-8 0.2807×10-2±3.22×10-8 0.1634×10-2±0.51×10-8 0.1157×10-2±0.95×10-8 0.1427×10-2±0.43×10-8 MSE 0.0613×10-2±3.26×10-9 0.0649×10-2±3.87×10-9 0.0322×10-2±1.04×10-9 0.0042×10-2±0.04×10-9 0.0052×10-2±0.06×10-9 0.0040×10-2±0.01×10-9 PSNR 31.6845±0.25 31.4407±0.26 34.4763±0.25 43.3787±0.34 42.3866±0.33 43.5526±0.13 SSIM 0.9667±6.22×10-6 0.9793±4.38×10-6 0.9753±1.89×10-6 0.9948±0.10×10-6 0.9940±0.22×10-6 0.9947±0.14×10-6
Table 4. Quantitative scores of soft tissue-based material decomposition results with different methods unit: mean±variance
View tableTable 4. Quantitative scores of soft tissue-based material decomposition results with different methods unit: mean±variance
Method Matrix Inversion Iterative Decomposition FCN Butterfly-net DIWGAN MC-UCTransNet MAE 2.3850×10-2±6.30×10-7 2.2024×10-2±5.85×10-7 0.6133×10-2±0.43×10-7 0.4009×10-2±0.17×10-7 0.3295×10-2±0.24×10-7 0.3466×10-2±0.16×10-7 MSE 0.2148×10-2±1.96×10-8 0.2202×10-2±1.98×10-8 0.0313×10-2±0.04×10-8 0.0115×10-2±0.08×10-8 0.0125×10-2±0.02×10-8 0.0112×10-2±0.03×10-8 PSNR 24.5465±0.08 24.4386±0.08 32.9122±0.08 36.3110±0.58 36.9219±0.22 37.3843±0.37 SSIM 0.8732±2.19×10-5 0.9317±0.47×10-5 0.9539±0.23×10-5 0.9807±0.19×10-5 0.9814±0.11×10-5 0.9815±0.15×10-5
Table 5. Quantitative scores of decomposition results of bone-based material images under different models unit: mean±variance
View tableTable 5. Quantitative scores of decomposition results of bone-based material images under different models unit: mean±variance
Model （w/o）Attention （w/o）SSIM Max-Min normalization MC-UCTransNet MAE 0.1271×10-2±1.09×10-8 0.1772×10-2±1.56×10-8 0.0933×10-2±0.65×10-8 0.0924×10-2±0.59×10-8 MSE 0.0129×10-2±3.09×10-10 0.0123×10-2±2.51×10-10 0.0085×10-2±1.55×10-10 0.0098×10-2±2.17×10-10 PSNR /dB 38.3890±0.37 38.5899±0.34 35.4132±0.50 39.5849±0.46 SSIM 0.9912±9.33×10-7 0.9917±8.66×10-7 0.9932±8.35×10-7 0.9939±6.53×10-7
Table 6. Quantitative scores of decomposition results of soft tissue-based material image under different models unit: mean±variance
View tableTable 6. Quantitative scores of decomposition results of soft tissue-based material image under different models unit: mean±variance
Model （w/o）Attention （w/o）SSIM Max-Min normalization MC-UCTransNet MAE 0.3258×10-2±0.13×10-7 0.4766×10-2±1.11×10-7 0.2369×10-2±0.07×10-7 0.2650×10-2±0.04×10-7 MSE 0.0159×10-2±3.10×10-10 0.0278×10-2±6.91×10-10 0.0069×10-2±0.41×10-10 0.0061×10-2±0.27×10-10 PSNR /dB 35.9114±0.22 33.4747±0.17 29.9182±0.16 39.9546±0.13 SSIM 0.9832±0.91×10-6 0.9826±0.92×10-6 0.9859±2.02×10-6 0.9904±0.15×10-6
Table 7. Quantitative comparison of model decomposition results under different scale feature combinations
View tableTable 7. Quantitative comparison of model decomposition results under different scale feature combinations
Model Bone Soft tissue PSNR /dB SSIM PSNR /dB SSIM 39.5849 0.9939 39.9546 0.9904 38.0509 0.9922 35.6849 0.9857 39.4329 0.9934 38.8624 0.9892 （w/o）Attention 38.3890 0.9912 35.9114 0.9832
Table 8. Comparison of the complexity of different methods
View tableTable 8. Comparison of the complexity of different methods
Method Number of parameters FLOP /109 Train time /min Test time /s FCN 49 448 92 24 Butterfly-net 3 1702 106 20 DIWGAN 19 1905 128 43 MC-UCTransNet 67 354 94 110
Tools
Get Citation
Copy Citation Text
Fan Wu, Tong Jin, Guorui Zhan, Jingjing Xie, Jin Liu, Yikun Zhang. Dual-Energy CT Base Material Decomposition Method Based on Multi-Channel Cross-Convolution UCTransNet[J]. Acta Optica Sinica, 2024, 44(5): 0515001
Paper Information
Category: Machine Vision
Received: Oct. 31, 2023
Accepted: Dec. 14, 2023
Published Online: Mar. 19, 2024
The Author Email: Liu Jin (liujin@ahpu.edu.cn)
CSTR:32393.14.AOS231715
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20