Enhancing CBCT-based CT synthesis using planning MRI in adaptive proton therapy for head cancer: A deep learning...

Proton therapy (PT) is recognized as a superior treatment for head cancer (HC) due to its precision and minimal damage to surrounding healthy tissues, relying on computed tomography (CT) data for dose calculations. Adaptive proton therapy (APT) is crucial to address changes in patient anatomy during treatment and update dose accuracy. However, in-room cone-beam CT (CBCT) assistance is limited to assessing patient setup, with occasional constraints due to artifacts and/or lower image quality and resolution compared to a CT scan.

Although deep learning (DL) techniques can successfully convert a CBCT into a synthetic CT (sCT), soft tissue delineation remains a challenging task. We hypothesized that, by including Magnetic Resonance Image (MRI) in CBCT-based CT synthesis, the sCT generation could more closely approximate the CT ground truth while improving tissue definition and dose calculation in PT treatment planning.

We propose a Pix2Pix-conditional generative adversarial network (cGAN) to synthesize a CT scan by combining two different input images: CBCT and T1-weighted MRI. ResUnet and SwinUnet were evaluated as the cGAN generator. Additionally, a CBCT-only-based CycleGAN was tested.

Model performance improved with the inclusion of MRI data, especially in recovering soft tissue details like eyes and ventricles, with ResUnet models outperforming SwinUnet models. Our cGAN outperformed both the self-autoencoder approaches and the CycleGAN model. Pix2Pix-ResUnet (MR-based) significantly reduced average HU errors in volumes of interest and also enhanced the precision in dose values, as demonstrated in dose differences and profiles.

We demonstrated the promising contribution of MRI to CBCT-based CT synthesis, enhancing sCT image quality and dose calculation accuracy. Future efforts should aim to collect a larger dataset and validate the integration of MRI in APT.