Deep learning in computed tomography super resolution using multi-modality data training.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Limited annotated data hinders AI in X-ray imaging.
• CT and X-ray share imaging physics, enabling cross-domain data sharing.
• Super-resolution (SR) can enhance CT resolution for generating realistic digital reconstructed radiographs (DRRs).

Purpose of the Study:

• Propose a novel SR network trained with kernel-based low-resolution (LR) and high-resolution (HR) images.
• Generate realistic multi-detector CT (MDCT) like LR images from HR cone-beam CT (CBCT) scans.
• Improve upon bicubic interpolation for SR in medical imaging.

Main Methods:

• Utilized a SR U-Net architecture for LR-HR mapping on CBCT image slices.
• Trained two models: SRUN (kernel-based LR) and SRUN (bicubic downsampled baseline).
• Evaluated models on unseen CBCT and MDCT datasets.

Main Results:

• Both SRUN models showed significant improvements in MAE, PSNR, and SSIM on unseen CBCT images.
• SRUN (kernel-based) outperformed SRUN (bicubic) with 14% MAE reduction, 6% PSNR, and 8% SSIM increase.
• SRUN generated sharper images and demonstrated cross-modality improvements on LR MDCT data.

Conclusions:

• The proposed SR method surpasses traditional interpolation for unseen LR CBCT images.
• Data frequency characteristics are crucial for learning SR features.
• The approach enables high-resolution CT-generated DRRs for deep learning training.