Computed tomography synthesis from magnetic resonance imaging using cycle Generative Adversarial Networks with...

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate dose calculation in MRI-only radiotherapy requires synthetic CT (sCT) generation from MRI.
• Deep learning (DL) shows promise for sCT synthesis, but methods often lack generalizability due to being center-specific.
• Multicenter training is a proposed solution to improve the generalizability of sCT synthesis models.

Purpose of the Study:

• To develop and evaluate a generalizable multicenter approach for DL-based sCT synthesis.
• To assess the impact of multicenter training on sCT accuracy and dose calculation in prostate cancer patients.

Main Methods:

• A 2D cycle-GAN was used for sCT synthesis from MRI data of 128 prostate cancer patients across four centers.
• Four training/testing scenarios were compared: monocenter, monocenter transfer, multicenter with unseen center test, and multicenter with seen center test.
• sCT accuracy was evaluated using Mean Absolute Error, Mean Error, and Peak-Signal-to-Noise-Ratio; dose accuracy was assessed via gamma index and Dose Volume Histogram comparison.

Main Results:

• Monocenter training and multicenter training with a seen center in the test set yielded comparable sCT accuracy.
• When tested on an unseen center, sCT quality significantly degraded compared to monocenter approaches.
• Inclusion of even limited data from one center in multicenter training improved generalization.

Conclusions:

• Generalizable multicenter training is crucial for robust DL-based sCT synthesis in MRI-only radiotherapy.
• Incorporating data from a single center within a multicenter training cohort can achieve sCT accuracy equivalent to monocenter studies.
• This approach enhances the applicability of sCT synthesis across different institutions.