Development and validation of machine learning-based rapid plan model for radiotherapy treatment of esophageal cancer
Sanju Sanju, Ajay Choubey, Torsha Chakraborty, Alka Kataria, Sanjay Barman, Vinay Saini, Narender Kumar, Sambit S. Nanda, Ashutosh Mukherji, Satyajit PradhanABSTRACT
Introduction:
Knowledge-based planning (KBP) enables high-quality radiotherapy plans with less time and variability. This study evaluated the clinical utility of RapidPlan® (Varian KBP system) in routine esophageal cancer treatment planning.
Materials and Methods:
A retrospective dataset of 75 treatment plans from patients with locally or locally advanced esophageal carcinoma was used to configure and train KBP models, while 20 additional plans were reserved for validation. Considering the phased treatment approach, two KBP models were created: Phase 1 delivered 45 Gy in 25 fractions to the planning target volume (PTV), and Phase 2 delivered a 9 Gy boost in five fractions to the PTV boost. Model quality was evaluated using the coefficient of determination (R 2 ), Chi-square (χ 2 ), and mean square error (MSE). For validation, RapidPlan® (RP) was compared with manual plans (MP) and RapidPlan with manual intervention (RPM), assessing dose-volume histograms (DVH), homogeneity index (HI), Paddick conformity index (PCI), and organ-at-risk (OAR) doses.
Results:
HI and PCI values were comparable across RP, MP, and RPM, confirming consistent target coverage. DVH analysis showed similar heart and spinal cord doses in all approaches. For the lungs, RP yielded higher low-dose volumes than MP (V10: 37 Gy vs. 34 Gy; V5: 57 Gy vs. 54 Gy). Manual refinement (RPM) enhanced lung sparing, reducing V10 to 33 Gy and V5 to 52 Gy.
Conclusion:
KBP with RP demonstrated efficiency and consistency, while manual adjustments further optimized OAR sparing, particularly for the lungs, offering a clinically feasible, time-saving alternative to conventional planning.