Deep nets vs expert designed features in medical physics: An IMRT QA case study.

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Intensity Modulated Radiation Therapy (IMRT) Quality Assurance (QA) is crucial for patient safety.
• Predicting gamma passing rates is essential for IMRT QA.
• Traditional methods rely on domain expert-designed features.

Purpose of the Study:

• To compare Deep Neural Networks (DNNs) with expert-designed models for predicting IMRT QA gamma passing rates.
• To evaluate the performance of Convolutional Neural Networks (CNNs) in this prediction task.

Main Methods:

• Trained CNNs using fluence maps from 498 IMRT plans.
• Employed TensorFlow, Keras, VGG-16 inspired architectures, and data augmentation.
• Compared CNN performance against a generalized Poisson regression model with 78 expert features.

Main Results:

• DNNs achieved performance comparable to domain expert models.
• An ensemble of neural networks yielded a Mean Absolute Error (MAE) of 0.70 ± 0.05.
• The domain expert model resulted in an MAE of 0.74 ± 0.06.

Conclusions:

• CNNs with transfer learning can predict IMRT QA passing rates by automatically extracting features.
• AI-driven predictions are comparable to those from expert-designed systems.
• This approach reduces the need for human expert supervision in feature engineering.