Online Adaptive Proton Therapy - Project Summary/Abstract: Intensity-modulated proton therapy (IMPT) has distinct advantages in terms of high conformality of target coverage and organs-at-risk (OAR) protection. However, IMPT is also subject to greater sensitivity to inter-fractional anatomical changes, which is commonly observed in head and neck (H&N) cancers and detrimental to the patient outcomes of IMPT. Adaptive radiation therapy (ART) has been introduced to account for the inter-fractional anatomical changes by introducing at least one re-plan during treatment courses. However, the gain of ART is at the cost of increased clinical workload. Online ART places greater demands not only on high efficiencies for each component of ART but also on their integration. Furthermore, the frequency of re-planning is significantly higher in proton therapy than photon therapy, making the same clinical tasks even more labor-intensive in proton clinics, particularly for H&N cancer. We propose to apply artificial intelligence (AI) to every ART component to enable the clinical use of online ART in IMPT. An asymmetric autoencoder network built with Vision-Transformer (ViT) blocks, which can learn the relationship between pairs of 2D kV images and 3D CTs. The synthesized 3D CTs solely based on 2D orthogonal kV images has better soft tissue details and can significantly improve patient alignment accuracy in some proton center without 3D on-board imaging (OBI) capability. The same technique can benefit photon therapy in low- income, rural areas, where many photon machines lack 3D OBI capability. Contour-aware ViT models will be developed to achieve real time deformable image registration (DIR) for contour propagation and dose deformation. We will further leverage contours and deformable vector fields (DVFs) generated during previous treatment planning as support dataset to achieve patient-specific adaptive DIR in online ART. Regarding re- planning, the reported AI-based dose calculation or influence matrix (IM) calculation methods are unary in the sense that a highly specialized network is trained for a single disease site. We will develop a universal AI model inspired by the latent diffusion model and hierarchical Bayesian. The proposed model will integrate knowledge from multiple data modalities of multiple disease cites in a single framework and yield efficient and effective IMs denoising for multiple disease sites with high accuracy. Since this method combines the first principle-physics (i.e., Monte Carlo) and AI (i.e., denoising) and thus it is more reliable and explainable. This work proposes a comprehensive approach to enable the clinical use of online adaptive proton therapy in IMPT to treat H&N cancer. We have proposed the state-of-the-arts AI tools to help expedite the whole workflow of online adaptive proton therapy. Success of this project will produce a next-generation treatment planning system (TPS) for translational research in IMPT, leading to significantly improved clinical outcomes. Furthermore, the success of this project can transform the standard of care for IMPT in treating other cancers, where significant inter-fractional anatomical changes are observed.
