INR‑CBCT: An Implicit Neural Representation Framework Enabling High‑Quality Sparse‑View Cone‑Beam CT for Image‑Guided Radiotherapy
Abstract
Purpose
In this preliminary study, we introduce an iterative implicit neural representation method for improving sparse‑view CBCT (INR-CBCT) for image-guided-radiotherapy.
Methods
INR-CBCT uses multi-resolution hashes with adaptive raw sampling to map coordinates to attenuation, serving as a custom regularization within a model based iterative reconstruction (MBIR) framework. A Catphan phantom was scanned on a Hitachi dual‑source CBCT. Projections were uniformly subsampled to 10, 25, 50, 100, 200, and 495 and reconstructed with FDK, MBIR, and INR-CBCT. Geometric and HU accuracy, and spatial resolution (MTF50) were calculated. For a small pilot study, projections from nine retrospective head-and–neck (HN) patients were collected and used to create scans; TotalSegmentator was used to make HN contours. Dice similarity coefficient (DSC) and mean-surface-distance (MSD) were computed comparing contours on reconstructed CBCTs vs clinical CBCTs.
Results
INR-CBCT outperforms all methods in PSNR for phantom and patient reconstructions at 100 projections and above, achieving a signal quality increase of up to 11.9% over FDK. Furthermore, it maintains superior structural accuracy (SSIM) at low projection counts in 78 seconds on average. HU accuracy was within acceptable clinical baseline limits for all Catphan materials and passed for all sparsely reconstructed scans. For Catphan reconstruction using >=50 projections, INR-CBCT had the lowest noise out of all methods. For >=50 projections, INR-CBCT maintained MTF50 within 0.1 lp/mm of the FDK-495 reference. Geometric distortion was within close range (0.10-0.29 mm) of the FDK--495 reference across all projections. INR-CBCT contours compared closely with clinical CBCT contours and were 0.703+/-0.349 DSC and 8.832+/-18.621mm MSD for 10-50 projections and 0.889+/-0.274 DSC and 2.924+/-9.013 MSD for 100 projections.
Conclusion
INR-CBCT is a new technique that enables aggressive projection downsampling while maintaining geometric accuracy, HU fidelity, and spatial resolution. Auto‑segmentation performance is sustained near clinical baseline at 50–100 projections.