Prospective Study of the Impact of Artificial Intelligence Organ-at-risk Dose Prediction on Head and Neck Radiotherapy Plan Quality

Abstract

Radiation treatment planning for head and neck cancer patients is complex, as the target and normal tissue anatomy are intertwined. The lowest achievable dose for each individual organ-at-risk (OAR) is unknown a priori. Artificial intelligence (AI) holds promise as an accurate predictor of the expected dose distribution for OARs. We prospectively studied a decision-support tool (DST) that leverages AI to predict OAR doses by identifying closely matched patients previously treated in an institutional database. The DST used an AI-based dose prediction model based on 376 volumetric modulated arc therapy (VMAT) plans previously treated at our institution. Under an IRB-approved prospective trial, the physician first generated a custom OAR directive for 50 consecutive patients (physician directive, PD) without engaging the DST. The treating physician then used the DST to identify the OAR doses from the closest matched patient (AI directive, AD), and the final Hybrid Directive (HD) for each OAR was the lower value of the AD and PD. The HD OAR metrics were used as the goals for treatment planning. A dose difference of 3 Gray (Gy) was considered clinically significant. Out of the 50 patients, the AD and PD estimates were within 3 Gy for 55% of all OARs, ranging from 27% (esophagus) to 78% (inferior constrictor). The AD was significantly lower for 23% of all estimates (ranging from 13% for the superficial parotid gland to 39% for superior constrictor dose), whereas the PD was lower in 22% of all estimates (ranging from 6% for inferior constrictor to 31% of the contralateral submandibular gland). When one directive was more than 3Gy lower than its comparison, the clinical plan typically met the lower constraint and the achieved average OAR dose reductions of 8.0Gy (AD >3Gy low and used) and 10.9Gy (PD >3Gy low and used). The reduction of OAR doses is mostly statistically significant (P<0.05). Use of the DST improved the final plan metrics in a significant percentage of patients (P<0.05, range 0.0001-0.046), but strictly relying on its estimates would have led to an inferior plan in an equal number of cases. While a more robust model informed by additional cases may further improve the accuracy of AI-based prediction tools, physician input is currently still imperative to generate an optimal plan.Abstract 197; TableOARPredictionAchieved PlanAD is LOWER than PD by > 3 Gy (%case)PD is LOWER than AD by > 3 Gy (%case)PD & AD within 3 Gy (%case)AD is LOWER than PD by > 3 GyPD is LOWER than AD by > 3 Gy%case AchievedAvg OAR reduction (Gy)%case AchievedAvg OAR reduction (Gy)Esophagus26%47%27%83%10.3*100%11.2*Larynx21%18%62%100%7.2*83%6.1*Constrictor_Sup39%17%44%100%7.5*100%13.6*Constrictor_Mid27%27%46%100%11.5*100%12.8*Constrictor_Inf16%6%78%60%8.1*50%6.6*Parotid_CL16%16%68%29%5.580%9.2*Sup_Parotid_IL13%10%77%100%6.9*100%12.3*SMG_CL15%31%54%100%7.2*100%19.4*Oral Cavity33%26%41%60%8.3*100%6.4**Statistical significance of the dose comparison, P < .05 Open table in a new tab