Deep Learning Projection Interpolation for Sparsely Sampled Real Patient CBCT Reconstruction
K Lu1*, Z Zhang1, L Ren2, F Yin1,3, (1) Duke University, Durham, NC, (2) University of Maryland, Baltimore, MD, (3) Duke University Medical Center, Durham, NC
Presentations
SU-H330-IePD-F5-2 (Sunday, 7/10/2022) 3:30 PM - 4:00 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 5
Purpose: Previous studies have used deep learning (DL) techniques for sinogram interpolation before CT reconstruction and achieved encouraging results with reduced patient dose. However, few studies have applied DL methods on real patient projection interpolation for cone-beam (CBCT) reconstruction. This study develops a DL-based model that performs projection interpolation on sparsely sampled data before CBCT reconstruction and post-processes the reconstructed images for improved image quality and reduced patient imaging dose.
Methods: Since real patient CBCT projection angles are not exactly evenly spread, the sparsely sampled projections are linearly interpolated according to densely sampled angles. The proposed technique re-slices a stack of linearly interpolated CBCT projections axially. Each acquired slice is input into a deep residual U-Net (DRU) model. The slices processed by DRU are reassembled to acquire a stack of processed densely-sampled projections, from which the CBCT volume is reconstructed with the FDK algorithm. A U-Net-based model further post-processes the reconstructed CBCT volume to improve the image quality. The proposed technique was compared with conventional linear interpolation on sparsely-sampled real patient CBCT projection data (98 extracted from 680 projections). A quantitative analysis was conducted with metrics of peak-signal-to-noise-ratio (PSNR), structural-similarity-index-measure (SSIM), and root-mean-square-error (RMSE).
Results: The PSNR, SSIM, and RMSE values for the sparsely-sampled FDK-reconstructed volume were 20.32, 0.68, and 0.10, respectively. For the CBCT volume reconstructed from linearly interpolated projections, the values were 24.30, 0.78, and 0.06. The scores for DRU reconstructed volume were 28.98, 0.84, and 0.04. After U-Net-based post-processing, the corresponding values were 29.84, 0.86, and 0.03. The proposed DL-based interpolation and post-processing technique showed an overall improvement in terms of PSNR, SSIM, and RMSE.
Conclusion: The proposed technique can perform DL-based interpolation on sparsely-sampled real patient CBCT projections and post-process reconstructed volumes to achieve improved image quality with reduced patient imaging dose.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by Duke University Chancellor Scholarship, the National Institutes of Health under Grant No. R01-EB028324, R01-EB001838, R01-CA226899, and a research grant from Varian Medical Systems.
Keywords
Cone-beam CT, Computer Vision, Data Interpolation
Taxonomy
IM/TH- Cone Beam CT: Machine learning, computer vision
Contact Email