Uncertainty-guided man-machine integrated patient-specific quality assurance.

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Radiation Oncology

Background:

• AI-based automatic patient-specific quality assurance (pQA) models enhance efficiency in radiation therapy.
• Current models lack uncertainty quantification, limiting safe clinical translation.
• Predicting gamma passing rate (GPR) is crucial for pQA.

Purpose of the Study:

• To develop an uncertainty-guided man-machine integrated pQA (UgMi-pQA) method.
• To address the limitation of state-of-the-art automatic pQA models in quantifying prediction uncertainty.
• To enhance the safety and clinical applicability of AI-based automatic pQA models.

Main Methods:

• Developed an uncertainty-aware dual-task deep learning (UDDL) model.
• Employed interwoven training and Monte Carlo dropout approximation Bayesian inference.
• Used 1541 GPR and fluence pairs for training/validation and 413 for out-of-distribution (OOD) detection.

Main Results:

• Achieved 100.0% clinical accuracy with 61.7% workload reduction.
• Successfully screened out samples with prediction errors and low GPR (<95%).
• Demonstrated significantly higher prediction uncertainty for OOD samples compared to in-distribution samples (p < 0.01).

Conclusions:

• Presents the first uncertainty quantification for deep learning automatic pQA.
• The UgMi-pQA method balances efficiency and safety of automatic pQA models.
• Facilitates the clinical application of AI in radiation therapy quality assurance.