Dosimetric evaluations using cycle-consistent generative adversarial network synthetic CT for MR-guided adaptive...

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Magnetic resonance (MR) guided radiation therapy (MRgRT) offers high-resolution imaging but lacks electron density data for accurate dose calculation.
• This limitation hinders the full application of MR-guided treatments.

Purpose of the Study:

• To evaluate Deep Learning (DL) based synthetic CT (sCT) generation from 3D MRI scans for adaptive MRgRT.
• To assess the accuracy of DL-generated sCT for radiation dose calculations.

Main Methods:

• A Cycle-consistent Generative Adversarial Network (Cycle-GAN) was trained using paired MRI and CT (dCT) data from various tumor sites.
• The DL model generated sCT volumes from MR setup scans.
• sCT accuracy was evaluated using image quality metrics and dosimetric calculations.

Main Results:

• The DL model achieved a Mean Absolute Error (MAE) of 49.2±13.2 HU between sCT and dCT.
• sCT demonstrated improved structural similarity and comparable electron density to dCT.
• Dosimetric evaluations showed minimal differences between sCT and dCT, with superior reconstruction of anatomical details like air-bubbles.

Conclusions:

• DL-based sCT generation from MR setup scans is accurate enough for dose calculation in MR-Linacs.
• This facilitates MR-only treatment planning workflows, enhancing efficiency in MRgRT simulation and adaptive planning.