Abstract:

Radiation-induced toxicity (RIT) poses a significant challenge in radiotherapy (RT). The goal of this study is to develop a framework for personalized RT treatment planning that leverages patient-specific data and dosimetric information to design an optimization model restricting adverse side effects via constraints that are learned from historical data.
This study employs the optimization with constraint learning (OCL) framework to directly incorporate patient-specific factors into the optimization process. The OCL framework consists of three main steps: first, the baseline treatment plan is optimized using population-wide dosimetric constraints; second, a machine learning (ML) predictive model is trained to estimate the patient's RIT to the baseline plan; third, the treatment plan is adapted to minimize RIT using ML-learned patient-specific constraints. We showcase our approach using several predictive models including classification trees, ensembles of trees, and neural networks, to predict the probability of grade 2+ radiation pneumonitis (RP2+) for non-small cell lung (NSCLC) cancer patients three months post-RT. The methodology is evaluated using four NSCLC patients with a high risk of RP2+ with the goal of optimizing the dose distribution to constrain the RP2+ outcome below a pre-specified threshold.