Improvement of IMRT QA prediction using imaging-based neural architecture search

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Quality assurance (QA) in intensity-modulated radiation therapy (IMRT) is crucial for patient safety.
• Predicting gamma passing rate (GPR) using machine learning (ML) has shown promise.
• Traditional methods often rely on hand-designed deep learning architectures.

Purpose of the Study:

• To apply a novel neural architecture search for optimizing deep neural networks (DNNs) in IMRT QA.
• To automatically tune and identify the best DNN architectures for predicting GPR.
• To compare the performance of automated DNNs against traditional ML models.

Main Methods:

• Developed 182 IMRT plans, delivered and measured using portal dosimetry.
• Utilized 1497 fields from multiple treatment sites as input for a convolutional neural network (CNN).
• Employed Auto-Keras with ResNet and DenseNet base models for CNN architecture search and fluence image regression.

Main Results:

• The deep learning approach achieved 98.3% prediction accuracy within 3% of measured 2%/2-mm GPRs.
• Maximum error was 3.1%, with a mean absolute error below 1%.
• Performance was comparable to ML approaches using handcrafted features.

Conclusions:

• A novel CNN model with an imaging-based neural architecture was successfully implemented for IMRT QA prediction.
• This deep learning method automates network architecture selection, eliminating manual feature extraction.
• The approach offers an efficient and accurate alternative for IMRT QA analysis.