Supervised versus unsupervised GAN for pseudo-CT synthesis in brain MR-guided radiotherapy

Area of Science:

• Medical imaging and artificial intelligence
• Radiotherapy treatment planning
• Deep learning for medical image synthesis

Background:

• Radiotherapy is essential for brain tumor treatment.
• CT-based planning has limitations; MR-only radiotherapy is an emerging alternative.
• Accurate MR-to-CT synthesis is crucial for MR-only radiotherapy.

Purpose of the Study:

• To compare supervised (Pix2Pix) and unsupervised (CycleGAN) deep learning for MR-to-CT synthesis.
• To evaluate the quality and clinical feasibility of pseudo-CT (pCT) images generated from MRI.
• To assess the performance of T1- and T2-weighted MR sequences for pCT generation.

Main Methods:

• Trained Pix2Pix (supervised) and CycleGAN (unsupervised) models on 3270 paired MRI and CT images.
• Generated pCT images from T1- and T2-weighted MR sequences.
• Assessed image quality using SSIM, PSNR, and MAE; performed dosimetric evaluation.

Main Results:

• Pix2Pix significantly outperformed CycleGAN in generating high-fidelity pCT images (p < 0.05).
• Average SSIM, PSNR, and MAE for Pix2Pix on T1 images were 0.964±0.03, 32.812±5.21, and 79.681±9.52 HU.
• No significant difference in pCT generation effectiveness between T1 and T2 MR sequences (p > 0.05); dosimetric evaluation showed comparable dose distributions.

Conclusions:

• Both supervised and unsupervised methods can generate accurate pCT images from conventional MR sequences.
• Supervised Pix2Pix offers higher accuracy, while unsupervised CycleGAN provides flexibility for unpaired data scenarios.
• MR-only radiotherapy planning using synthesized pCT images is clinically feasible.