Synthetic CT generation from CBCT and MRI using StarGAN in the Pelvic Region

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate dose calculation in radiotherapy requires computed tomography (CT) data.
• Magnetic resonance imaging (MRI) and cone-beam CT (CBCT) are increasingly used in radiation therapy, but lack direct CT Hounsfield unit (HU) information.
• Synthetic CT (sCT) generation aims to bridge this gap, enabling dose calculation from MRI/CBCT.

Purpose of the Study:

• To evaluate StarGAN, a deep learning model for generating sCT from both MRI and CBCT data.
• To compare StarGAN's performance against CycleGAN for sCT generation.
• To assess the dosimetric accuracy and anatomical preservation of sCT images for radiotherapy applications.

Main Methods:

• Utilized StarGAN and CycleGAN models for sCT generation.
• Dataset included 53 pelvic cancer cases.
• Evaluated sCT quality using qualitative and quantitative analyses, including dose distribution calculations.

Main Results:

• StarGAN demonstrated superior anatomical preservation in sCT generated from CBCT and MRI.
• CycleGAN showed lower mean absolute error (MAE) for body and bone HU values.
• Both models achieved acceptable dosimetric accuracy (mean dose difference <2%, gamma passing rate >90%).

Conclusions:

• StarGAN excels in anatomical accuracy for sCT generation, crucial for radiotherapy.
• Despite higher quantitative errors, StarGAN's anatomical fidelity presents significant potential for MRI simulation and adaptive radiation therapy.
• The choice between StarGAN and CycleGAN may depend on the specific requirements for quantitative accuracy versus anatomical detail.