Deep learning-based multi-modality image conversion for evaluating dose calculation and position correction accuracy in image-guided radiation therapy.

Cone-beam CT (CBCT) is widely used in image-guided radiation therapy (IGRT). However, CBCT has limited image quality, which can reduce dose calculation and position correction accuracy. This study aimed to propose a deep learning-based multi-modality image synthesis framework to enhance the feasibility of adaptive radiation therapy (ART) and IGRT. Two conditional generative adversarial network (CGAN) models were developed: a CBCT-to-CT model to generate synthetic CT (sCT) from CBCT, and a CT-to-MRI model to generate synthetic MRI (sMRI) from CT. The models were trained with 80 cases and tested with another 20 cases. Image quality was evaluated using the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). Moreover, a volumetric modulation arc therapy plan of 48 Gy in four fractions was recalculated using sCT images and compared with the dose distributions of original CT images. Each sMRI and original MRI were registered to the original CT, and displacements between each registered image were quantified based on the CT. The sCT generated using the CBCT-to-CT model improved SSIM from 0.15 to 0.82 and PSNR from 11.3 to 19.9 dB, while the sMRI generated using the CT-to-MRI model improved SSIM from 0.09 to 0.63 and PSNR from 10.3 to 23.9 dB. The dose distribution based on sCT achieved a gamma pass rate over 95%, and positional correction differences of sMRI were within 1 mm. Multi-modality image synthesis using CGAN enables high-fidelity image generation from CBCT, improving both dose calculation and position correction accuracy for ART and IGRT.