Self-supervised CT super-resolution with hybrid model.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Software-based methods offer a way to improve Computed Tomography (CT) spatial resolution.
• Current methods may require hardware modifications or increased radiation exposure.
• Deep learning (DL) presents a promising avenue for enhancing CT image quality.

Purpose of the Study:

• To develop a DL-based CT super-resolution (SR) method for reconstructing low-resolution (LR) sinograms into high-resolution (HR) CT images.
• To integrate CT domain knowledge into a hybrid SR and deblurring model.
• To create a self-supervised DL network trainable with minimal data.

Main Methods:

• Developed a hybrid model (SADIR) combining sinogram-domain SR and image-domain deblurring.
• Unrolled the SADIR model into a DL network (SADIR-Net) incorporating CT imaging principles.
• Employed a self-supervised learning approach for SADIR-Net, enabling training and testing with single sinograms.

Main Results:

• SADIR-Net achieved superior performance on physical and porcine phantoms compared to state-of-the-art DL methods.
• Demonstrated significant improvements in Information Fidelity Criterion (IFC), Structural Similarity Index (SSIM), and Root-Mean-Square Error (RMSE).
• Showcased substantial enhancements in Modulation Transfer Function (MTF), improving MTF50% by 69.2% and MTF10% by 69.5% over Filtered Back Projection (FBP).

Conclusions:

• SADIR-Net effectively reconstructs HR CT images from LR sinograms, offering performance comparable to existing methods.
• The method excels in scenarios with limited or no training data.
• SADIR-Net provides a viable solution for improving CT spatial resolution without altering scanner hardware or increasing radiation dose.