BLUE RIBBON POSTER THERAPY: Devlopment of a Decision Support Tool for Prioritizing Breast RT Cases In Peer Review Rounds
Abstract
Purpose
Peer review is a critical quality assurance step in radiation therapy (RT), but not all cases require the same level of attention. We aimed to develop a machine learning (ML) tool to help prioritize breast RT plans for peer review based on their likelihood of prompting discussion or requiring modification before treatment.
Methods
We analyzed 1424 breast-only, locally advanced, and partial-breast irradiation RT plans presented in peer review. We labelled each plan based on whether it elicited discussion among the peer-review group and/or required changes [1] or if it was unanimously approved without deliberation [0]. A random forest classifier was trained per cohort using over 100 handcrafted features—including dose-volume metrics, shape descriptors, and radiomic features from dose distributions. Additional flags for known planning challenges (e.g., surgical clip coverage and breast folds) were incorporated. Feature selection was informed by statistical tests and clinical input.
Results
The random forest classifier demonstrated strong discrimination across all cohorts. For breast-only and locally advanced cases, left-sided plans had slightly higher performance (AUC = 0.865) than right-sided (AUC = 0.845), reflecting anatomical complexity near the heart. Partial breast plans, analyzed together due to limited sample size, also yielded robust performance. Lowering the classification threshold increased recall from 0.70 to 0.82 and reduced precision (0.92 to 0.71), favouring sensitivity to flag critical cases. The most influential features included radiomic dose metrics within the planning target volume and anatomical distances such as heart-to-target.
Conclusion
This ML tool can effectively flag breast RT plans more likely to require peer review attention. Stronger performance for left breast cases may reflect the added complexity of planning near the heart. The tool shows promise in supporting more focused, efficient peer review and safer treatment planning. These results will guide future fine-tuning to enhance model sensitivity and specificity.