Multiple Kernel Synthesis of Head CT Using a Task-Based Loss Function.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Head CT scans often involve multiple reconstructions with varying kernels and slice thicknesses.
• Reviewing redundant information from these reconstructions is inefficient for radiologists.

Purpose of the Study:

• To develop a convolutional neural network (CNN)-based technique, ZIRCON, to create a single, high-quality CT image series.
• To combine favorable features from smooth and sharp head kernels into one optimized image.

Main Methods:

• Utilized a CNN with an autoencoder U-Net architecture accepting smooth- and sharp-kernel CT images as input.
• Employed a task-based loss function with region-specific smooth and sharp loss terms.
• Trained the model using supervised learning with routine-dose clinical images as targets and simulated low-dose images as inputs.

Main Results:

• ZIRCON produced thinner slices and improved gray-white matter contrast, especially with the smooth-kernel loss function.
• Reduced noise and improved visibility of small soft-tissue features, mitigating issues from partial volume averaging and noise.
• Line profile analysis indicated ZIRCON images largely retained sharpness compared to sharp-kernel inputs.

Conclusions:

• ZIRCON effectively combines desirable image quality properties from both smooth and sharp kernels into a single, thin, low-noise series.
• The technique is suitable for both brain and skull imaging, enhancing diagnostic capabilities.
• ZIRCON offers an efficient solution for reviewing head CT data, improving workflow for radiologists.