A convolutional neural network model for EPID-based non-transit dosimetry.

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Imaging

Background:

• Accurate dose verification is crucial in Intensity-Modulated Radiation Therapy (IMRT).
• Electronic Portal Image Devices (EPIDs) are commonly used for in-vivo dosimetry.
• Existing methods for non-transit dosimetry using EPID data have limitations.

Purpose of the Study:

• To develop an alternative computational approach for EPID-based non-transit dosimetry.
• To utilize a convolutional neural network (CNN) model for dose distribution prediction.

Main Methods:

• A U-net CNN architecture with a non-trainable layer was developed.
• The model was trained on 186 IMRT Step & Shot beams to convert portal images to absolute dose distributions.
• Model performance was evaluated using the gamma index and compared against an existing algorithm.

Main Results:

• The developed CNN model achieved high accuracy, with average gamma-passing rates of 99.29% for clinical beams.
• The model demonstrated superior performance compared to an existing analytical portal image-to-dose conversion algorithm.
• Sufficient model accuracy was achievable with the amount of training data utilized.

Conclusions:

• A deep learning-based model effectively converts portal images into absolute dose distributions.
• This novel method shows significant potential for enhancing EPID-based non-transit dosimetry in radiation therapy.