Adversarial Prediction of Radiotherapy Treatment Machine Parameters

TL;DR

This paper introduces a generative adversarial network-based model that predicts radiotherapy treatment machine parameters, aiming to streamline plan creation and improve initial plan quality, especially in sparing organs-at-risk.

Contribution

It develops a novel cGAN-based approach to predict treatment machine parameters, serving as a lower bound and initialization for plan refinement, enhancing efficiency and quality in radiotherapy planning.

Findings

Predicted plans achieve comparable OAR sparing to clinical plans.

Initial plans require refinement for target coverage to match clinical quality.

Predicted plans can significantly reduce computation time for plan optimization.

Abstract

Modern external beam cancer radiotherapy applies prescribed radiation doses to tumor targets while minimally affecting nearby vulnerable organs-at-risk (OARs). Creating a treatment plan is difficult and time-consuming with no guarantee of optimality. Knowledge-based planning (KBP) mitigates this uncertainty by guiding planning with probabilistic models based on populations of prior clinical-quality plans. We have developed a KBP-inspired planning model that predicts plans as realizations of the treatment machine parameters. These are tuples of linear accelerator (Linac) gantry angles, multi-leaf collimator (MLC) apertures that shape the beam, and aperture-intensity weights that can be represented graphically in a coordinate frame isomorphic with projections (beam's-eye views) of the patient's target anatomy. These paired data train conditional generative adversarial networks (cGANs) that estimate the MLC apertures and weights for a novel patient, thereby predicting a treatment plan. The predicted plans' OAR sparing is close to that of the clinical plans; the predicted target coverage requires refinement to match the clinical plans' quality. Nonetheless, the predicted plans can serve as lower bounds on plan quality, and by initializing the MLC aperture shape and weight refinement can substantially reduce the compute times for that refinement.