Streamlining Thoracic Radiotherapy Quality assurance: One-Class Classification for Automated OAR Contour Assessment.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Automated contouring in radiotherapy treatment planning reduces physician workload and improves consistency.
• Manual quality assurance (QA) for automated contours is time-consuming and subjective.
• An effective automated QA process for organ-at-risk (OAR) delineations is currently needed.

Purpose of the Study:

• To develop and evaluate an automated QA model for OAR contours generated by auto-segmentation algorithms.
• To assess the model's ability to detect various contour quality levels and types of errors.
• To reduce the burden on physicians in radiotherapy planning.

Main Methods:

• Utilized the AAPM Thoracic Auto-Segmentation Challenge dataset for OARs (lungs, heart, esophagus, spinal cord).
• Employed a ResNet-152 network as a feature extractor and a one-class support vector machine (OC-SVM) for classification.
• Generated low-quality contours via translation/resizing and assessed correlations with metrics like Dice Similarity Coefficient (DSC) and Hausdorff Distance 95% (HD95).

Main Results:

• The OC-SVM model demonstrated superior performance over binary classifiers in balanced accuracy and Area Under the ROC Curve (AUC).
• The model effectively detected diverse contour errors with high interpretability.
• Strong correlations were found between automated detection limits and established contour metrics (volume, DSC, HD95, MSD).

Conclusions:

• An attention-mechanism integrated OC-SVM framework automates QA for OAR delineations in radiotherapy.
• This approach accurately detects various contour errors, significantly easing the workload for physicians.
• The model offers a robust solution for ensuring contour quality in automated radiotherapy planning.