Author: Anthony J. Doemer, Bing Luo, Benjamin Movsas, Humza Nusrat, Farzan Siddiqui, Chadd Smith, Kundan S Thind, Kyle Verdecchia 👨🔬
Affiliation: Department of Physics, Toronto Metropolitan University, Henry Ford Health 🌍
Purpose: Large-vision models (LVMs) are rapidly emerging, yet their application in radiation oncology remains largely unexplored. This study investigates the potential of LVMs for offline adaptive radiotherapy in head-and-neck (HN) cancer. We propose a framework that utilizes a LVM to analyze anatomical changes from daily cone-beam computed-tomography (CBCT) images, including delta couch shifts, to flag offline adaptive treatment planning in HN cancer patients.
Methods: First, a patient dataset query was performed to highlight HN patients with previous deformable image registration, since these patients likely represent offline adaptive radiation therapy within our clinic. The pretreatment localizing CBCTs were retrieved, including the applied delta couch shifts during treatment setup. For each CBCT, a post-processing mask was applied to each slice identifying pixels as either air or non-air. Furthermore, all pixels on the masked slices were summed creating a weighted 2D image that represents all non-air pixels (>-500 HU) from an individual CBCT; transforming the 3D CBCT to a derived 2D image required for the LVM. Lastly, a relative comparison between derived 2D images for the initial treatment (i.e. baseline) and each ensuing treatment fraction created a differential 2D imageset per fraction.
Results: The previously mentioned data were retrieved for 62 patients. An illustrative timeline of the differential 2D imageset visually highlights the relative anatomical changes as patient treatment progressed. Further work will investigate the use of these generated anatomical images with a LVM to determine the feasibility of flagging adaptive radiotherapy via artificial intelligence.
Conclusion: The retrospective treatment course data-retrieval and data-processing performed in this work provides the initial framework for utilizing a LVM to trigger adaption for HN patients. Anatomical changes during treatment course are illustrated as a 2D image and will be used to assess the feasibility of artificial intelligence to guide on-treatment clinical decisions within radiation oncology.