Author: Leigh A. Conroy, Thomas G Purdie, Christy Wong 👨🔬
Affiliation: Department of Medical Biophysics, University of Toronto, Princess Margaret Cancer Centre 🌍
Purpose: To develop a novel machine learning (ML) algorithm to evaluate and rank breast radiation therapy (RT) treatment plans based on treatment complexity for prioritization in multidisciplinary peer review rounds.
Methods: We analyzed over 1000 breast only and locally advanced breast RT plans that were presented in peer review rounds. Each RT plan was classified as complex or non-complex. Complex RT plans either elicited discussion among the peer review group and/or identified the requirement for updating the RT plan. A random forest classifier was trained for each cohort using >100 handcrafted features (e.g., dose-volume histograms, geometric shape metrics, radiomic properties, spatial distances across segmented anatomy). To address dataset imbalance, additional complex RT plans were manually created, and models were trained using an 80/20 train/test split.
Results: For breast only RT plans, the model achieved an AUC of 0.84 for both left and right breast cases. For locally advanced breast RT plans, performance was higher for left breast cases (AUC 0.94) compared to right breast cases (AUC 0.85). Histogram features for the target and ipsilateral lung, as well as spatial distances between targets and organs at risk, were key predictors. Decision threshold adjustments to prioritize sensitivity ensured more complex RT plans were identified, achieving an average recall of 0.81 and precision of 0.68 across the dataset.
Conclusion: The developed model can accurately identify complex breast RT plans. The superior performance for left breast plans may reflect increased complexity due to proximity of targets to critical structures such as the heart, providing the model with more distinct features for classification. These results will guide future fine-tuning of the ML model to further enhance sensitivity and specificity. This work demonstrates the potential of ML to automate and streamline case prioritization in RT peer review rounds for efficient and accurate clinical decision-making.