Author: Yunxiang Li, Xinlong Zhang, You Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Acquiring high-resolution (HR) proton density (PD) images is time-consuming, while lower-resolution (LR) PD scans are faster but can lack sufficient details. We propose CycleHR, a T2-contrast-guided, unpaired super-resolution approach to generate HR PD from LR scans at multiple down-sampling factors, aiming to maintain tissue fidelity without relying on paired LR–HR datasets.
Methods: CycleHR trains on unpaired HR PD and T2 images, eliminating the need for matched LR–HR PD. The network takes an LR PD (with HR T2 guidance) as input, learns to synthesize HR PD through adversarial and cycle-consistency losses, and is benchmarked against classical upsampling methods (cubic, Lagrange, and BM3D). We evaluated structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and learned perceptual image patch similarity (LPIPS) on multiple test PD images. At a small undersampling factor (s=2), we expected classical interpolation methods to perform well. At larger factors (s=4, s=8), we hypothesized that CycleHR would better preserve structural details due to T2 guidance and adversarial training.
Results: At s=2, cubic, Lagrange, and BM3D often matched or exceeded CycleHR numerically, reflecting easier upsampling tasks. However, at s=4 and s=8, these methods experienced substantial drops in both SSIM (<0.80) and PSNR (~21–23 dB), alongside higher (worse) LPIPS scores (≥0.05). In contrast, CycleHR maintained SSIMs of ~0.83–0.86 and PSNRs of ~25 dB at s=8, indicating improved structural fidelity. LPIPS values (~0.04–0.05) further confirmed its perceptual advantage over standard interpolation approaches for more challenging upsampling scenarios.
Conclusion: By leveraging unpaired HR PD and T2 data, CycleHR addresses the challenge of lengthy HR PD acquisition durations, enabling super-resolution on clinically-convenient LR scans to recover crucial structural details. This T2-guided framework offers robust super-resolution performance to improve PD image quality without requiring paired LR-HR PD training data, particularly benefiting higher undersampling scenarios where traditional interpolation methods struggle.