Characterization of a Bayesian network-based radiotherapy plan verification model.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Current radiotherapy treatment plan quality assurance is manual, time-consuming, and prone to errors.
• A Bayesian network (BN) was previously proposed to assist in this process.
• This study expands and trains the BN to better reflect clinical practice.

Purpose of the Study:

• To expand and train a Bayesian network (BN) for improved radiotherapy treatment plan quality assurance.
• To evaluate the BN's performance using historical clinical data and simulated errors.
• To determine optimal training data parameters for the BN.

Main Methods:

• Utilized 51,540 radiotherapy cases (2010-2017) from an Elekta oncology information system.
• Developed a 29-node, 40-edge BN based on clinical experience and machine learning (expectation maximization).
• Tested the BN with withheld data, introduced errors, and varying training dataset sizes, lengths, and eras, evaluating performance via Area Under the receiver operating characteristic Curve (AUC).

Main Results:

• Achieved AUCs of 0.82, 0.85, 0.89, and 0.88 with 2-yr, 3-yr, 4-yr, and 5-yr training windows, respectively.
• A 4-yr sliding window showed a 3% AUC reduction per year when shifted back in time.
• Loss of detection performance in plan/beam errors contributed most to AUC reduction over time.

Conclusions:

• The expanded BN effectively detects common radiotherapy planning errors.
• A 4-year training dataset optimizes BN performance for this institutional data.
• Yearly BN updates are sufficient to adapt to evolving clinical practice and maintain accuracy.