Sparse-View 4D-CBCT Enhancement Via Deep Learning with Integrated Interpolation and Motion Compensation
Abstract
Purpose
Four-dimensional cone-beam computed tomography (4D-CBCT) enables visualization of respiration-induced anatomical motion but suffers from degraded image quality due to sparse-view acquisition. This study proposes a deep learning-based approach combining interpolation and motion compensation (Interpolation+MoCo) to enhance sparse-view 4D-CBCT.
Methods
The Interpolation+MoCo consists of two stages: (a) Interpolation: an interpolation network that predicts full-view projections from sparse views using a U-Net backbone with a self-designed, angular-aware convolution that adaptively modulates kernel weights based on projection rotation angles and assigns diverse filters to different input patches. (b) MoCo: a deformable image registration (DIR) network that warps images from other phases to the target phase to mitigate undersampling artifacts. A total of 100 full-view 4D-CBCT projection sets from 20 patients were split into training (12 patients, 60 sets), validation (4 patients, 20 sets), and test (4 patients, 20 sets) cohorts. Each full-view set (Φfull) was retrospectively downsampled from 2500 to 900 projections to generate sparse-view data (Φsparse). Network training consisted of two stages. (a) Interpolation network: the projections in Φfull-Φsparse served as targets, each paired with its two adjacent projections in Φsparse as inputs, yielding 90190 and 30245 training and validation pairs respectively. (b) DIR network: trained in an unsupervised manner using source-target image pairs reconstructed as three-dimensional (3D) CBCT from the interpolated projection sets, with 5400 and 200 training and validation pairs respectively. The Interpolation+MoCo was tested on 20 4D-CBCT sets using root-mean-square error (RMSE), structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR).
Results
The proposed Interpolation+MoCo approach significantly enhance sparse-view 4D-CBCT quality with a RMSE of (3.25±0.51)×10-3 mm-1, a SSIM of 0.9905±0.0061 and a PSNR of 35.09±1.65 dB, outperforming sparse-view+MoCo results (RMSE: (4.79±0.84)×10-3 mm-1, SSIM: 0.9887±0.0055, PSNR: 31.73±1.10 dB).
Conclusion
The proposed Interpolation+MoCo approach enables high-quality 4D-CBCT from sparse views with shortened scanning-time and reduced dose.