A comparison of knowledge-based dose prediction approaches to assessing head and neck radiotherapy plan quality

Abstract

PURPOSE: Recent studies demonstrate deep learning dose prediction algorithms may produce results like those of traditional knowledge-based planning tools. In this exploratory study, we compared 2D DVH-based knowledge-based planning tools and METHODS: Pre-validated 2D and 3D dose prediction models were applied to 58 patients with head and neck cancer treated under RTOG 0522 obtained from The Cancer Imaging Archive (TCIA). The 2D model was used to predict dose-volume histogram bands for seven organs at risk (OARs; brainstem, spinal cord, oral cavity, larynx, mandible, right parotid, and left parotid). A 3D dose prediction model was used to predict 3D dose distributions, based on computed tomography images, OAR contours, planning target volumes and prescriptions. The mean and D1% to the seven OARs for the 2D and 3D dose prediction models were compared. Further post predictive analysis was done to quantify the predicted 3D dose sparing for all normal tissues. RESULTS: The two models predicted similar dose sparing to the OARs, with a mean difference of 1.4± 5.5 Gy across all evaluated dose metrics. When looking at the sparing of non-OAR normal tissue regions, the 3D model predicted a mean dose reduction to normal tissue regions of 6.4±3.0 Gy when compared with the clinical dose. CONCLUSION: 2D and 3D dose predictions are comparable at predicting dose reductions to OARs. The 3D approach allows for dose visualization, which may support further sparing of normal tissues not typically drawn as OARs on head and neck plans.