2287 - Feasibility Study of an AI-Assisted MRI-Only Simulation-Free Radiation Therapy Workflow
Presenter(s)
Z. Yang1, G. A. Szalkowski2, E. Rahimy1, J. Lewis1, D. Pham3, Y. Yang3, Y. Gao3, L. Wang4, S. G. Soltys1, K. Zhang5, M. Kazemimoghadam6, Q. Wang6, M. Chen6, H. Jiang7, W. Lu6, and X. Gu8; 1Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, 2Georgia Institute of Technology, Atlanta, GA, 3Department of Radiation Oncology, Stanford University, Stanford, CA, 4Department of Radiation Oncology, Stanford University School of Medicine, Palo Alto, CA, 5Department of Radiation Oncology, Stanford University, s, CA, 6Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX, 7Department of Radiation Oncology, The University of Texas Southwestern Medical Center, Dallas, TX, 8Stanford University Department of Radiation Oncology, Palo Alto, CA
Purpose/Objective(s): Radiotherapy is a common and effective treatment regimen for cancer care, with expedited workflows being critical for symptom management, particularly in palliative care. However, the reliance on CT simulation can create bottlenecks, leading to treatment delays. This study aims to develop an AI-assisted, MRI-only, simulation-free workflow and evaluate its feasibility on palliative brain radiotherapy.
Materials/Methods: The workflow was developed and implemented on a web-based in-house image analysis platform clinically used at our institution. We retrospectively collected 120 patients with T1 contrast-enhanced MRI and paired simulation CT (sim-CT) available to validate the workflow. The proposed workflow consists of five steps: (1) MRI is imported into the platform, where an in-house AI model segments critical brain structures (e.g., brainstem, eyes, lens). (2) Another AI model is also implemented to generate a synthetic CT from the imported MRI, with the auto-segmentations propagated to the synthetic CT accordingly. (3) The platform auto-registers the synthetic CT to a template CT with standard head positioning using rigid registration to improve setup reproducibility. (4) The last slice of the synthetic CT is duplicated to extend the inferior field of view (=5 cm below the brain) to ensure accurate scatter modeling in dose calculation. (5) The platform sends the synthetic CT and auto-contours to the treatment planning system, where 3D bilateral whole-brain plans are then generated using an automated script. To validate the accuracy and quality of the synthetic CT and treatment plans, we randomly selected 10 patients for quantitative analysis. Hounsfield unit (HU) differences between the synthetic CT and sim-CT were evaluated for skull and brain tissue. Additionally, dosimetric comparisons were performed by recalculating 3D plans using sim-CTs.
Results: The HU differences between the synthetic CT and sim-CT were 54±27.3 for skull and 28±14.7 for brain tissue. Dosimetric comparisons showed deviations of <1% for D95% and D99% of the PTV between the plans calculated with the synthetic CT and sim-CT. The differences in maximum dose and V105% of the PTV were 1.3%±0.6% and 11%±9%, respectively.
Conclusion: Preliminary results demonstrate the feasibility of an AI-powered, simulation-free radiation therapy workflow using pre-existing MRI images. It achieved clinically acceptable synthetic CT image quality and plan dosimetric accuracy. This workflow could provide potential benefit to both patients and the hospital by reducing costs and expediting time to treatment. Further evaluation on VMAT treatment is warranted to assess broader clinical applicability.