Semi-Automated Hippocampal Sparing Avoidance Whole-Brain Radiotherapy Using a Commercially Available Artificial Intelligence Toolkit

Abstract

Hippocampal-avoidance whole-brain radiotherapy (HA-WBRT) plus memantine has been proven to better preserve cognitive function in brain metastasis patients receiving WBRT. Recent results from the phase III trial NRG Oncology CC001 have shown HA-WBRT plus memantine can reduce cognitive deterioration without affecting intracranial progression free survival (PFS) or overall survival (OS), and should be considered standard of care for patients without metastases in the hippocampi. However, treatment planning techniques that are capable of reducing dose to the hippocampi typically involve inverse planning techniques (IMRT/VMAT) that require an extended planning timeline when compared to the prior standard of care using 2D planning. This timeline could be significantly reduced by leveraging available semi-automated planning algorithms. The purpose of this work was to evaluate the feasibility of using an artificial intelligence (AI)-assisted, commercially available, semi-automated planning algorithm for HA-WBRT. Six de-identified patients previously treated at our institution with HA-WBRT using VMAT were re-planned using a commercial, semi-automated treatment planning system workflow. Treatment plans were generated to deliver 30 Gy in 10 fractions, using target coverage and normal tissue constraints from NRG Oncology CC001. The clinical goals template was first iteratively optimized to generate plans that were comparable to the clinical VMAT plan and compliant with NRG CC001 using a single patient dataset. This same template was then used to automatically generate plans for the remaining 5 datasets. The semi-automated system generates five different plans for comparison that include a 5, 9 and 12-field IMRT plan, as well as a 2 and 3-arc VMAT plan. The automatically generated treatment plans were reviewed and compared with the previously delivered clinical VMAT plans. The automatically generated 9-field IMRT field consistently respected all NRG CC001 constraints and was chosen as the best plan for all six datasets. All constraints (or their acceptable variations) were met for all six datasets without any intervention. PTV_3000 coverage was comparable or slightly improved compared to clinical plans. On average, V30Gy for PTV_3000 was 0.7% higher (93.8% vs 93.1%) and D98% was 0.9 Gy higher (25.9 Gy vs 25.0 Gy), while D2% was reduced by 1.2 Gy (33.0 Gy vs 34.2 Gy). The dose to the hippocampi was increased but still within NRG CC001 constraints, with D0.03cc 1.6 Gy higher (14.6 Gy vs 12.9 Gy) and D100% 0.4 Gy higher (8.5 Gy vs 8.1 Gy). Dose to the optic nerves and optic chiasm was consistently reduced by an average of 2.5 Gy (31.9 Gy vs 29.3 Gy). A commercially available, semi-automated workflow aided by AI can generate comparable HA-WBRT plans as traditional inverse planning techniques while reducing the planning time. Further optimization can improve plan quality and reduce planning time to aid the widespread adoption of HA-WBRT.