Computed tomography super-resolution using deep convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) imaging often involves trade-offs between slice thickness, resolution, and noise.
• Improving the resolution and reducing noise in CT images is crucial for accurate diagnosis.

Purpose of the Study:

• To develop a convolutional neural network (CNN) for CT image super-resolution.
• To enable an end-to-end learning of the mapping between low-resolution (thick-slice) and high-resolution (thin-slice) CT images.

Main Methods:

• A modified U-Net architecture was employed as the CNN.
• The network was trained and tested using 2D slices from CT studies, with axially averaged thick-slice data as input and corresponding middle slices as ground truth.
• Five-fold cross-validation was utilized to ensure performance consistency.

Main Results:

• The CNN generated super-resolved images virtually equivalent to the ground truth.
• Significant improvements were observed in deblurring bone structure and air cavity boundaries.
• The CNN output showed a ~10% higher peak signal-to-noise ratio and lower normalized root mean square error compared to the input thick-slice images.
• Noise levels in the CNN output were lower than the ground truth and comparable to iterative reconstruction methods.

Conclusions:

• The developed deep learning method effectively performs super-resolution on CT images.
• The CNN also demonstrates significant de-noising capabilities.
• This approach offers a valuable tool for enhancing CT image quality, potentially improving diagnostic accuracy and reducing radiation dose.