Error detection model developed using a multi-task convolutional neural network in patient-specific quality assurance...

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning

Background:

• Machine learning (ML) methods for dose analysis in intensity-modulated radiation therapy (IMRT) QA can identify error sources.
• These ML methods have not been widely applied to volumetric-modulated arc therapy (VMAT) QA.
• Detecting errors in VMAT QA is crucial for patient safety and treatment efficacy.

Purpose of the Study:

• To propose and evaluate a deep learning approach for detecting various errors in patient-specific VMAT QA.
• To develop a multi-task convolutional neural network (CNN) capable of individually identifying 12 distinct error types.

Main Methods:

• Analyzed 161 prostate VMAT beams using a cylindrical detector (Delta4).
• Simulated 12 types of errors (MLC position, MU scaling, gantry rotation, phantom setup) and created 13 dose difference maps.
• Trained a multi-task CNN model on these maps and evaluated its performance against gamma analysis on two test datasets.

Main Results:

• The CNN model achieved higher accuracy than gamma analysis in classifying single-type errors on Test set 1 (0.92 vs. 0.81).
• On Test set 2, the CNN model's accuracy was 0.44, while gamma analysis with 2%/1mm criteria achieved 0.95.
• The model demonstrated superior performance in identifying individual error types compared to gamma analysis, though it was less effective with compound errors.

Conclusions:

• A multi-task CNN model was successfully developed for detecting errors in patient-specific VMAT QA.
• The developed model proved effective in identifying specific error types within VMAT QA dose maps.
• Further research may be needed to improve performance for compound errors.