Author: Michael Bowers, Patrik Brodin, Madhur Garg, Rafi Kabarriti, William P. Martin, Todd R. McNutt, Julie Shade, Wolfgang A. TomΓ©, Christian Velten π¨βπ¬
Affiliation: Johns Hopkins University, Oncospace, Inc., Montefiore Medical Center π
Purpose: Development of an automated planning tool utilizing AI generated patient-specific dose-volume histogram predictions for rapid H&N plan generation.
Methods: Planning best-practices were developed, generalized, and implemented in C# using the Varian Eclipse API to automatically generate plans using the TPSβs dose calculation algorithms. Dose-volume optimization objectives for organs-at-risk (OAR) were obtained from a cloud-based AI dose-volume histogram (DVH) prediction tool (Oncospace, Baltimore, MD), following the export of CT images and radiotherapy structures and the selection of a H&N treatment protocol. The developed tool was tested on the 2023 AAMD planning challenge H&N case (4-level SIB: 63/60/57/54Gy in 33 fractions). It was additionally compared to 16 institutional plans developed by senior dosimetrists, where the prescription doses were 69.96/59.4/54.12 Gy in 33 fractions. Plans were generated in batch mode for both Varian TrueBeam and Halcyon, each with IMRT and VMAT.
Results: Times for plan generation (excluding protocol selection and setup) were as low as 2min for IMRT and up to 15min with VMAT using GPU-accelerated optimization and dose calculation. The automatically generated plan scored 132.4/150 total points on the 2023 AAMD plan challenge, above the median score of all submitted plans (130). It scored better than the majority of submitted challenge plans on PTV coverage and OAR sparing. Automatically generated plans for the institutional dataset achieved comparable target coverage, ΞD99%-95%β0.46 Gy (90%CI: [-1.8, 3.3] Gy). Target heterogeneity (ΞD2%) did not increase significantly (90%CI: [-2.4, 3.3]). Substantial reductions in OAR Dmean and Dmax were achieved in most plans: for example, for the parotids the average ΞDmean was β3.3Gy (90%CI: [β12.2,3.7] Gy), while the spinal cord Dmax was lower by 11.6 Gy on average (90%CI: [β18.4,-3.8] Gy).
Conclusion: Automated plan generation yielded high-quality treatment plans within a fraction of typical planning times, allowing for increased efficiency and standardization of the planning process.