Multi-granularity prior networks for uncertainty-informed patient-specific quality assurance

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Automated Patient-Specific Quality Assurance (PSQA) in radiation therapy is crucial for safety and resource efficiency.
• Current deep learning models for PSQA lack uncertainty quantification, limiting clinical trust.
• Predicting dose difference metrics like Gamma Passing Rate (GPR) and Dose Difference Prediction (DDP) is vital for treatment validation.

Purpose of the Study:

• To develop a Multi-granularity Uncertainty Quantification (MGUQ) framework for deep learning-based PSQA.
• To integrate uncertainty quantification into multi-task PSQA, specifically for GPR and DDP prediction.
• To enhance the trustworthiness and clinical applicability of deep learning in radiation therapy quality assurance.

Main Methods:

• Proposed a Bayesian framework (MGUQ) for quantifying uncertainties at multiple granularities in PSQA.
• Developed a multi-granularity loss function incorporating granularity-specific and coherence loss components.
• Introduced Multi-granularity Prior Networks, a dual-stream architecture for inferring DDP (t-distributions) and GPR (Gaussian distributions).

Main Results:

• Achieved a minimum Mean Absolute Error (MAE) loss of 0.864 with a 2%/3 mm criterion for GPR prediction.
• Demonstrated uncertainty visualization of dose difference.
• Attained 100% Clinical Accuracy (CA) with a reduced workload of 67.2%.

Conclusions:

• The MGUQ framework effectively quantifies uncertainties in deep learning-based PSQA.
• The proposed method enhances the trustworthiness of AI in radiation therapy quality assurance.
• This approach improves the safety, effectiveness, and clinical adoption of automated PSQA systems.