Enhance four-dimensional cone-beam computed tomography (4D-CBCT) from sparse view acquisitions using a novel deep learning model

Four-dimensional cone-beam computed tomography (4D-CBCT) enables time-resolved imaging of moving targets but requires long scan times and high radiation doses for sufficient projections per respiratory phase. This study proposes a deep learning-based projection interpolation model with data-aware convolutional design to enable sparse-view 4D-CBCT with reduced imaging time and dose. We developed a two-stage workflow, Interpolation + MoCo, to enhance sparse-view 4D-CBCT. The Interpolation stage employs dynamic region-aware and rotation-aware convolutional modules to predict missing projections by adjusting weights based on angular distance and spatial subregions. The MoCo stage performs motion compensation using a deep learning-based deformable image registration (DIR) model. Both models were trained and evaluated on 100 full-view scans from 20 lung cancer patients. Full-view data (∼2500 projections) were downsampled to 900 sparse projections, with remaining projections used as targets. The interpolation model was trained on 90,190 sample sets (one set = {two sparse projections, one target}), validated on 30,245, and tested on 30,084. For DIR, reconstructed volumes from interpolated and original sparse projections formed phase pairs (training: 5400; validation: 200; test: 200). Performance was assessed using root mean square error (RMSE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR) against motion-compensated full-view reconstructions. The workflow enables high-quality 4D-CBCT from 1-minute sparse acquisitions. Interpolation achieved SSIM: 0.9657 ± 0.0155, RMSE: (8.38 ± 0.98) × 10⁻ 4 mm⁻ 1 , PSNR: 28.44 ± 1.56 dB. After MoCo, results improved to SSIM: 0.9905 ± 0.0061, RMSE: (32.46 ± 5.14) × 10⁻ 4 mm⁻ 1 , PSNR: 35.09 ± 1.65 dB. Interpolation + MoCo enables accurate, motion-compensated 4D-CBCT reconstruction from sparse-view data, supporting faster, lower-dose imaging for radiotherapy motion management.