Synthesizing images from multiple kernels using a deep convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) image reconstruction involves a trade-off between noise and spatial resolution based on kernel selection.
• Clinical practice often necessitates generating multiple CT images with varying kernels, increasing computational and logistical burdens.
• Deep learning offers a potential solution to optimize CT image quality and streamline clinical workflows.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN) capable of synthesizing a single CT image from multiple kernel-reconstructed inputs.
• To achieve a synthesized image that combines low noise levels with high spatial resolution, mimicking the best attributes of different reconstruction kernels.
• To address the clinical challenges associated with producing and managing multiple CT image series.

Main Methods:

• A 20-layer deep CNN architecture utilizing residual units was employed to combine image features.
• The CNN input comprised two CT images (smooth and sharp kernels) stacked channel-wise.
• Supervised learning was used for training with full-dose and simulated quarter-dose abdominal CT data.

Main Results:

• Synthesized images demonstrated noise levels comparable to or slightly lower than the smooth input images.
• Anatomic details in the synthesized images retained the high spatial resolution characteristic of the sharp input images.
• Quantitative evaluation confirmed noise reduction and resolution preservation through RMS and line profile analyses.

Conclusions:

• Deep CNNs can effectively integrate features from differently reconstructed CT images into a single, high-quality series.
• This synthesized approach yields images with both reduced noise and enhanced spatial resolution.
• The method holds promise for improving CT image quality, potentially lowering radiation dose, and simplifying clinical CT interpretation.