Purpose: To propose a novel hybrid approach, which utilizes prior planning CT (pCT) with deep learning, for correcting daily pretreatment CBCT to facilitate online verification of proton range.
Methods: Previously proposed methods for correcting CBCT using CycleGAN are often limited in suppressing image artifacts, while maintaining a high geometric fidelity. In contrast, classic methods of deforming pCT onto CBCT provide more accurate intensities but suffers from geometric uncertainties. In the proposed method, high and low spatial-frequency components in each CBCT and pCT image are separated via Fourier transform. Then, CycleGAN and deformation are respectively used to correct geometry-weighted (high frequency) and intensity-weighted (low frequency) components which are subsequently combined, thereby each process deals with only the component weighted toward its strength. The CycleGAN was trained with 50 high-frequency image pairs of abdominopelvic CBCT/CT. A fast deformation algorithm followed by a threshold-based internal air correction was employed to process pCT. Additional datasets of 4 and 7 CBCT/pCT image pairs were used respectively to tune meta-parameters and evaluate the mean absolute error (MAE) performance of the proposed method against the same-day CT.
Results: Correcting a CBCT dataset took less than a minute including the CycleGAN and deformation processes. The proposed method produced more accurate intensities than using CycleGAN alone (MAE, 37±4 vs. 42±4; P<0.001, paired t-test) and significantly reduced artifacts. When comparing with the deformation-alone method, the improved accuracy was pronounced in bony structures (defined by HU > 200; MAE, 73±12 vs. 87±19; P=0.045, paired t-test) reflecting robustness against misregistered bones in the deformed pCT.
Conclusion: In the proposed hybrid approach, CycleGAN and deformable registration complemented each other to perform better than either individual method. The improvement in bony structure fidelity has implications for estimating proton range which is sensitive to geometric errors in those structures.