A Deep Learning Technique for Projection Interpolation in CBCT Reconstruction
K Lu1*, Z Zhang1, L Ren2, F Yin2, (1) Duke University, Durham, NC, (2) Duke University Medical Center, Durham, NC
Presentations
TU-C-TRACK 3-1 (Tuesday, 7/27/2021) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Purpose: Some previous studies used DL techniques for sinogram interpolation prior to CT reconstruction to reduce the imaging dose. However, few studies have applied DL techniques on CBCT projection interpolation for cone-beam CT (CBCT) reconstruction. This study develops a novel DL technique that performs CBCT projection interpolation before the reconstruction step to improve CBCT image quality, reduce artifacts caused by sparse sampling, as well as reduce the imaging dose.
Methods: Due to the limitation of GPU memory, the proposed technique re-slice the stack of projection data along the axial direction. The processed data are fed into the custom-designed network slice by slice. As the network iterates through each slice in the input stack of sparsely sampled projections, the outputs of the network are stacked together to form the stack of fully sampled projections. These projections are then used to reconstruct CBCT with the Feldkamp, Davis, and Kress (FDK) algorithm. The proposed technique is tested for CBCT reconstruction using sparsely sampled data. We compared the CBCT images reconstructed from projections interpolated with the conventional bilinear interpolation method and the proposed technique to demonstrate its feasibility and advantages. The quantitative evaluation metrics used are peak-signal-to-noise ratio (PSNR), structural similarity (SSIM) index, and root-mean-square error (RMSE).
Results: Average PSNR, SSIM, and RMSE values are 41.552, 0.988, and 0.008 for CBCT images reconstructed from projections interpolated with the proposed DL technique; 37.774, 0.975, and 0.013 for bilinear interpolation method; 39.692, 0.978, and 0.010 for images reconstructed without projection interpolation. The proposed technique shows an overall improvement in PSNR, SSIM, and RMSE.
Conclusion: The proposed DL interpolation technique can perform accurate CBCT projection interpolation that reduces artifacts for CBCT reconstructed from sparsely sampled projections as well as the imaging dose.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by Duke University Chancellor Scholarship and is partially supported by NIH 1R01EB028324-01 and R01CA184173.
Handouts
Keywords
Cone-beam CT, Image Processing, Reconstruction
Taxonomy
IM/TH- Cone Beam CT: Machine learning, computer vision
Contact Email