Room: AAPM ePoster Library
Purpose: To perform a dosimetric comparison study comparing clinical whole brain radiotherapy plans and those created a deep learning approach for clinical application.
Methods: First, 312 clinical whole brain cases with blocks drawn by physicians were used to train a deep learning model to automatically define field apertures on laterally opposed beams using patient’s digitally reconstructed radiographs. A fully-automated planning tool was then implemented using Python scripting and normal tissue atlas-based automated contouring in RayStation. Final testing was performed on 40 new test cases for which clinical plans were already available. We compared doses to the brain, cribriform plate, and lens for the deep learning (DL) plans and clinical plans (recalculated in RayStation). The prescription dose was 30Gy in 10 fractions and all plans were normalized 100 % to the marked isocenter point. The cribriform plate contour was later added by physicians specifically for this dosimetric comparison.
Results: Brain dose coverage (D_99% and D_95%) for the DL plans were 29.7 Gy ± 0.48 Gy and 30.3 Gy ± 0.34 Gy, respectively, this was comparable with the clinical plans within 1 %. D_1% for the brain was 2.2 % lower in DL plans than the clinical plans. Cribriform plate dose (D_99%) was 17.1 % (3.3 Gy) lower than clinical plans and an average dose of both lenses was 18 % (0.7 Gy) lower than clinical plans.
Conclusion: Automated whole brain radiotherapy plans created using deep learning are dosimetrically comparable to clinical plans in regards to brain dose coverage and lower lens dose. The DL tool has been clinically deployed at our institution and prospective evaluation of this tool by physicians' discretion is underway.
Brain, Radiation Dosimetry, Treatment Planning
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation