Machine Learning to Predict the Quality of a Left-Sided Whole Breast Radiotherapy Treatment Plan

Abstract

<h3>Purpose/Objective(s)</h3> Develop a Support Vector Machine (SVM) model that uses dosimetric data to predict the quality of a breast radiotherapy treatment plan. <h3>Materials/Methods</h3> One hundred twenty left-sided breast cancer patients from a single institute were included in the study. They were treated with 15-fractions whole breast radiotherapy up to 40.05 Gy with a concomitant boost to deliver 48 Gy to the tumor bed. There was no treatment of lymph nodes. Dose was delivered with a fixed-beam treatment mode and delivered plans were manually created. For each patient, a second treatment plan was generated, using an in-house system for automated plan generation. The treating clinician evaluated the two available plans for each patient and assigned a sentence regarding the quality of the plan in the format of yes/no variable (well optimized plan i.e., "yes" or plan that can be further optimized i.e., "no"). A quantitative scoring tool (ST) already used by Orecchia et al was used to assign 0, 0.5 or 1 points for each involved OAR and PTVs depending on both the fulfilment of multiple DVH constraints and on achieved coverage level for PTVs. Four SVM models were trained using 216 plans (arbitrarily selected from the available 240) using: i) dose volume parameters for target and OARs of each plan, ii) ST and iii) the clinician's plan sentence. Training was performed through iterative optimization using ten Cross-validation folds with four models: a) Model 1 with the ST only, b) Model 2 with the difference between dose volume parameters and the DVH constraints, c) Model 3 with both information from models 1 and 2 and d) Model 4 with ST and dose volume parameter. All models had clinician's sentence as response variable. Finally, the 24/240 patients not used for training were used for final validation. Models' performances were assessed through confusion matrix parameters and multiclass ROC analyses on the training and independent validation set. <h3>Results</h3> The involved clinician refereed as "yes" 92 plans (77%) out of 120 and "no" the remaining 28 in the manual set; 79 plans (66%) were judged "yes" and 41 "no" in the automatic set. Results in terms of Accuracy (ACC), Positive Predicted Values (PPV), True Positive Rates (TPR) and Area under Curve (AUC) for the four models are presented in table 1. <h3>Conclusion</h3> Clinicians' judgements are crucial in daily plan evaluations. We have proposed supervised learning to predict if a plan will be considered by clinician well optimized or not. Results are encouraging and the tool could potentially become useful for aiding treatment planners and reducing clinicians' workload.