Development and Clinical Validation of a Protocol-Agnostic Machine Learning Platform for Automated Treatment Planning in External Beam Radiation Therapy.

PURPOSE: To develop and validate a protocol-agnostic machine learning platform ("Predictive Planning") for knowledge-based planning (KBP) in external beam radiation therapy. METHODS AND MATERIALS: Five thousand three hundred thirty-four retrospective photon beam treatment plans (1145 Head and Neck [H&N], 1623 Thoracic, 781 Abdominal, 1785 Pelvis) from 2 institutions were used to develop and internally test general-purpose models for organ at risk (OAR) dose-volume histogram prediction. Models were deployed in a commercial system (Plan AI, Sun Nuclear Corporation), and 72 retrospective clinical plans (18 H&N, 20 Thoracic, 17 Abdominal, 17 Pelvis) were replanned using treatment planning system optimization objectives predicted by the models ("Predictive Plans"; PPs), without objective value changes. Dose metrics were compared for 25, 11, 8, and 7 OARs for H&N, Thoracic, Abdominal, and Pelvis plans, respectively, and for planning target volumes. RESULTS: In the plan comparison study, there were no OAR dose metrics for which PPs were statistically significantly greater than clinical plans. For H&N, PP mean dose (Gy) was significantly lower for Brain (5.8 vs 7.3, P = .0003), Brainstem (9.6 vs 13.8, P < .0001), Glottis (26.9 vs 31.9, P = .0017), OpticChiasm (7.3 vs 12.2, P < .0001), Parotid_L/R (19.0/20.8 vs 22.3/25.7, P < .0001/<.0001) and SpinalCord (10.1 vs 15.7, P < .0001). For Abdominal, PP mean dose was significantly lower for Heart (2.2 vs 3.4, P = .0039), Kidney_R (6.2 vs 8.6, P = .002), SpinalCanal (4.8 vs 6.7, P = .0004), and Stomach (9.1 vs 10.9, P = .002). For Pelvis, PP mean dose was significantly lower for Bladder (30.7 vs 33.7, P < .0001), Femur_Head_L/R (13.8/13.9 vs 17.0/16.5, P = .0026/.0046), PenileBulb (15.8 vs 21.9, P = .0039), and Rectum (28.1 vs 34.2, P = .0002). Mean planning target volume coverage was significantly higher for H&N and Thoracic, and equivalent for Abdominal and Pelvis. CONCLUSIONS: Predictive planning represents a shift in KBP from protocol-specific models trained with small, uniform data sets to protocol-agnostic, disease-site-specific models trained with large, heterogeneous data sets. Plans generated using optimization objectives predicted by the models had equivalent or superior dosimetric outcomes.