Patient-specific quality assurance failure prediction with deep tabular models

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Patient-specific quality assurance (PSQA) failures in radiotherapy can lead to treatment delays and increased staff burden.
• Predicting PSQA failures in advance is crucial for optimizing radiotherapy workflows.
• Current methods may lack the precision needed for reliable failure prediction.

Purpose of the Study:

• To develop a predictive model for radiotherapy PSQA failures using only multi-leaf collimator (MLC) leaf positions.
• To create an end-to-end differentiable map from MLC leaf positions to PSQA failure probability.
• To potentially aid in regularizing optimization algorithms and generating more robust treatment plans.

Main Methods:

• A retrospective dataset of 968 volumetric arc therapy patient plans was used.
• An attention-based neural network, FT-Transformer, was trained on beam-level MLC leaf positions.
• Model performance was evaluated for predicting gamma pass rates (regression) and PSQA failure (classification), compared against tree ensemble methods (CatBoost, XGBoost) and a mean-MLC-gap metric.

Main Results:

• FT-Transformer achieved a 1.44% Mean Absolute Error in gamma pass rate prediction, comparable to XGBoost (1.53%) and CatBoost (1.40%).
• In PSQA failure classification, FT-Transformer achieved an ROC AUC of 0.85, outperforming the mean-MLC-gap metric (0.72 ROC AUC).
• FT-Transformer, CatBoost, and XGBoost achieved an 80% true positive rate with a false positive rate under 20%.

Conclusions:

• Reliable PSQA failure predictors can be developed using only MLC leaf positions.
• The FT-Transformer model provides an end-to-end differentiable mapping from MLC positions to PSQA failure probability.
• This approach offers a novel way to improve radiotherapy treatment planning and quality assurance.