Multi-task autoencoder based classification-regression model for patient-specific VMAT QA.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning Applications

Background:

• Patient-specific quality assurance (PSQA) for volumetric modulated arc therapy (VMAT) is crucial for accurate treatment delivery but is resource-intensive.
• Machine learning shows promise for predicting VMAT PSQA results, yet performance, especially for plans with low gamma passing rates (GPRs), requires enhancement and clinical validation.
• Existing models struggle with the classification accuracy and sensitivity needed for unbalanced VMAT plan datasets.

Purpose of the Study:

• To develop and validate a novel multi-task machine learning model, autoencoder based classification-regression (ACLR), for VMAT PSQA.
• To improve the prediction accuracy and classification sensitivity of VMAT PSQA results, particularly for plans with low GPRs.
• To evaluate the generalization performance of the ACLR model through clinical validation (CV).

Main Methods:

• Developed a novel autoencoder based classification-regression (ACLR) model integrating classification and regression tasks.
• Employed balanced sampling techniques to address unbalanced VMAT plan data.
• Utilized 54 plan modulation and delivery metrics as inputs to predict PSQA GPRs, validated on 426 plans (technical validation) and 150 plans (clinical validation).

Main Results:

• The ACLR model demonstrated significantly improved prediction accuracy compared to the Poisson Lasso (PL) model, with lower absolute prediction errors (APE) across different gamma criteria.
• In technical validation, ACLR achieved APEs of 1.76% (3%/3 mm), 2.60% (3%/2 mm), and 4.66% (2%/2 mm), outperforming PL.
• The ACLR model achieved 100% sensitivity and 83% specificity at 3%/3 mm, showing high accuracy in classifying unbalanced VMAT QA results with no significant difference between technical and clinical validation.

Conclusions:

• The developed ACLR model offers a robust and accurate solution for VMAT PSQA, significantly enhancing prediction accuracy and classification sensitivity.
• The model's ability to handle unbalanced datasets and its strong performance in both technical and clinical validation suggest its clinical utility.
• ACLR can be readily applied for virtual VMAT QA, potentially reducing the resource intensity of current PSQA processes.