Robustness of deep learning-based denoising 4DCBCT methods.
Authors
Affiliations (3)
Affiliations (3)
- The Netherlands Cancer Institute - Antoni van Leeuwenhoek Hospital, Plesmanlaan 121, Amsterdam, 110, 1066 CX, Netherlands.
- VISLab, University of Amsterdam, Science Park 904, 1098XH AMSTERDAM, Amsterdam, 1000 GG, Netherlands.
- Antoni van Leeuwenhoek Hospital, Nederlands Kanker Instituut - Antoni van Leeuwenhoek Ziekenhuis, Plesmanlaan 121, 1066 CX Amsterdam, THE NETHERLANDS, Amsterdam, Noord Holland, 1066 CX, Netherlands.
Abstract
4D cone-beam computed tomography (4DCBCT) is a technique used to address respiratory motion in radiotherapy but is limited by significant view-aliasing artifacts. Recently, deep learning methods have been proposed to reduce view-aliasing. This study investigates the robustness and performance of these methods when variations in patient breathing period affect the noise pattern in the scans used as training data, which potentially affects the ability of the models to reduce view-aliasing.

Approach. We evaluated both a supervised and a self-supervised scalable deep learning methods. Using a dataset of 328 patients, scans were partitioned into cohorts according to breathing period, and independent training with cross-cohort validation was performed to assess robustness. We also evaluated the impact of a modern UNet-based architecture and data augmentation. To quantify view-aliasing without requiring ground truth, we introduced the Power Spectral Gradient (PSG), a metric measuring respiratory-phase streak consistency in the Radon domain.

Main results. Training and testing on matched breathing-period cohorts did not consistently improve performance, indicating that both methods are robust to clinically observed respiratory-pattern variations. Instead, increasing the overall dataset size led to the most noticeable improvements. For the cohort with long breathing periods, the supervised model trained on the combined dataset achieved a ~0.5 PSNR improvement over the cohort-specific model and a PSG improvement of ~0.7. While data augmentation improved general image quality (increasing PSNR by ~0.3), it did not reduce view-aliasing artifacts, yielding comparable or slightly worse PSG.

Significance. These findings support the reliability of deep learning-based 4DCBCT denoising across clinical breathing patterns. They suggest that large, diverse datasets should be prioritized over breathing-cohort-specific optimization. Finally, PSG provides an objective tool for artifact quantification and may support future development of more robust 4DCBCT enhancement techniques.