Robust Multi-Site Deep Learning-Based Dose Predictions for Volumetric Modulated Arc Therapy Using the MR-Linac
Presentations
MO-I430-BReP-F3-4 (Monday, 7/11/2022) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 3
Purpose: To propose a generic deep learning (DL)-based solution for multi-site dose calculations for Volumetric Modulated Arc Therapy (VMAT) using the MR-Linac.
Methods: Our network was trained on the dose of individual multi-leaf-collimator (MLC) segments using the DeepDose framework. This physics-based approach uses a set of ray-traced 3D inputs to encode features of the photon transport through a patient anatomy. Our dataset consisted of a total of 100 tumour patients (42 prostate, 24 rectal, 34 lung), previously treated in our clinic with VMAT. We initially converted the clinical plans to MR-Linac-deliverable VMAT plans and we furthermore added a 1.5T external magnetic field to our ground truth dose calculations. Additionally, the gantry and collimator angles of each training sample were randomized and the segment shapes were rigidly shifted within the MLC space to prevent potential beam angle configuration biases. The trained model was then used to predict the dose for a set of test cases. For the evaluation of our method, the predicted dose distributions were benchmarked against the ground truth ones using the 1%/1mm, 2%/2mm and 3%/3mm gamma criteria and the corresponding dose-volume-histograms were assessed.
Results: Our model predicted highly accurate dose distributions for 11 previously unseen VMAT patient cases, while also modelling the electron return effect caused by the external magnetic field. The average gamma pass rate scores reported were: 99.36±1.12%, 97.47±3.41% and 99.71±0.36% for prostate, rectal and lung patient test fractions respectively for the clinical 3%/3mm criterion.
Conclusion: We presented a framework for DL-based dose calculations for multiple tumour sites in radiation therapy. The robustness of our method was proven for a broad set of VMAT patient plans, converted to a generic MR-linac treatment environment. We believe that our DL-based dose engine has the potential to be used in a clinical setup for accurate dose calculations.
Keywords
Taxonomy
Contact Email