The Digital Press

This study introduces a deep learning method for assessing the quality of auto-generated radiotherapy contours, focusing on online adaptive radiotherapy. The approach uses Bayesian Ordinal Classification with uncertainty calibration to accurately predict contour quality without extensive manual labeling.

Why it matters: It reduces manual workload and speeds up quality assurance in radiotherapy, improving clinical efficiency.

The big picture: The method works well with no, limited, or extensive labels, adapting through surrogate labels, transfer learning, or direct supervision.

Stunning stat: Fine-tuning with only 30 labels and calibrating with 34 subjects achieves over 90% accuracy in quality prediction.

Quick takeaway: Calibrated uncertainty thresholds identify over 93% of contour qualities correctly, minimizing unnecessary manual reviews in 98% of cases.