Dose Characteristics of a Deep Learning Model for EPID-based In vivo Dosimetry.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Electronic portal imaging devices (EPID) capture dose information in images.
• This information can be converted into 2D dose maps for dosimetry.
• A CycleGAN-based model was developed for this 2D EPID dosimetry application.

Purpose of the Study:

• To develop and evaluate a CycleGAN-based model for 2D EPID dosimetry.
• To assess the dose characteristics and accuracy of the developed model.
• To investigate the impact of different normalization methods on model performance.

Main Methods:

• Measurements were performed on a linac with an EPID detector.
• Dose distributions were calculated using a treatment planning system as ground truth.
• CycleGAN models converted EPID images to 2D dose maps using two normalization methods.
• Gamma analysis and dose linearity were used for model evaluation.

Main Results:

• EPID dose characteristics showed high precision, with linearity observed beyond 12 cm phantom thickness.
• The CycleGAN model effectively transformed EPID images into planar dose maps.
• Normalization method II achieved a 97.9% mean pass rate in gamma analysis (3 mm, 3%), significantly outperforming method I (85.5%).

Conclusions:

• EPID is a valuable tool for capturing dose information that can be accurately converted to planar dose maps using CycleGAN models.
• The developed model shows potential for quality assurance in radiation therapy treatment plans.
• Choosing appropriate normalization methods is crucial for mitigating dose nonlinearity and improving accuracy.