Generating synthetic computed tomography for radiotherapy: SynthRAD2023 challenge report

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Accurate patient anatomy representation is vital for radiation therapy dose calculations.
• Conventional Computed Tomography (CT) provides essential electron density data but is not always feasible for daily adaptive radiotherapy.
• Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast but lacks electron density information, while Cone Beam CT (CBCT) has calibration limitations.

Purpose of the Study:

• To benchmark and compare various synthetic CT (sCT) generation techniques.
• To evaluate sCT methods using multi-center data for MRI-to-CT and CBCT-to-CT tasks.
• To assess the clinical applicability of sCT in adaptive radiotherapy workflows.

Main Methods:

• The SynthRAD2023 challenge utilized multi-center ground truth data from 1080 patients.
• Two tasks were defined: MRI-to-CT and CBCT-to-CT image synthesis.
• Evaluation involved image similarity metrics and dose-based assessments using photon and proton treatment plans.

Main Results:

• The challenge attracted 617 registrations with 22/17 valid submissions for tasks 1/2.
• Top methods achieved high structural similarity indices (≥0.87/0.90) and gamma pass rates (≥98.1%/99.0% for photons, ≥97.3%/97.0% for protons).
• No significant correlation was observed between image similarity and dose accuracy.

Conclusions:

• Deep learning demonstrates significant potential for generating high-quality sCT, reducing reliance on conventional CT.
• Dose evaluation is critical for assessing the clinical utility of sCT, beyond image quality metrics.
• SynthRAD2023 provides valuable insights for advancing MRI-only and CBCT-based adaptive radiotherapy.