Evaluating the Use of Machine Learning to Predict Expert-Driven Pareto-Navigated Calibrations for Personalised Automated Radiotherapy Planning

Abstract

Automated planning (AP) uses common protocols for all patients within a cancer site. This work investigated using machine learning to personalise AP protocols for fully individualised planning. A ‘Pareto guided automated planning’ (PGAP) solution was used to generate patient-specific AP protocols and gold standard Pareto navigated reference plans (MCOgs) for 40 prostate cancer patients. Anatomical features related to geometry were extracted and two ML approaches (clustering and regression) that predicted patient-specific planning goal weights were trained on patients 1–20. For validation, three plans were generated for patients 21–40 using a standard site-specific AP protocol based on averaged weights (PGAPstd) and patient-specific AP protocols generated via regression (PGAP-MLreg) and clustering (PGAP-MLclus). The three methods were compared to MCOgs in terms of weighting factors and plan dose metrics. Results demonstrated that at the population level PGAPstd, PGAP-MLreg and PGAP-MLclus provided excellent correspondence with MCOgs. Deviations were either not statistically significant (p ≥ 0.05), or of a small magnitude, with all coverage and hotspot dose metrics within 0.2 Gy of MCOgs and OAR metrics within 0.7% and 0.4 Gy for volume and dose metrics, respectively. When compared to PGAPstd, patient-specific protocols offered minimal advantage for this cancer site, with both approaches highly congruent with MCOgs.