Uncertainty quantification-guided patient-specific quality assurance using Bayesian neural networks based on field...

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiation Oncology

Background:

• Advancements in AI for patient-specific quality assurance (PSQA) require robust uncertainty quantification (UQ) for clinical safety.
• Current AI models for PSQA need methods to ensure reliability and guide clinical decision-making.

Purpose of the Study:

• To develop and validate an uncertainty-guided framework for AI-driven PSQA prediction.
• To improve the clinical safety and efficiency of automated PSQA systems.

Main Methods:

• An AI classification model was trained using field complexity and fluence maps to categorize PSQA outcomes.
• Monte Carlo approximate Bayesian inference was used for UQ, with thresholds defined by Correct-Certain (CC) and Incorrect-Uncertain (IU) curves.
• A Multilayer Perceptron predicted gamma passing rates (GPR), integrating classification embeddings and uncertainty metrics.

Main Results:

• The classification model demonstrated high sensitivity across different gamma criteria (e.g., 94.74% at 2%/2 mm).
• The framework reduced manual intervention by 42.73% while achieving 100% clinical sensitivity at 3%/3 mm.
• Integrating uncertainty improved GPR accuracy for 'failed' cases by 21.03% and prospective tests showed 100% accuracy, sensitivity, and specificity.

Conclusions:

• The proposed uncertainty-guided framework enhances the reliability of AI-based PSQA.
• This approach significantly reduces manual workload and improves prediction accuracy, facilitating the safe clinical adoption of automated PSQA.