Aliasing layers for processing parallel imaging and EPI ghost artifacts efficiently in convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance (MR) imaging techniques like parallel imaging and Echo Planar Imaging (EPI) can introduce aliasing and ghosting artifacts.
• Existing Convolutional Neural Network (CNN) methods struggle with these artifacts due to a lack of artifact structure consideration, requiring extensive layers.

Purpose of the Study:

• To develop an efficient method for processing MR-specific artifacts using CNNs.
• To introduce a novel layer that accounts for artifact structures for improved removal.
• To demonstrate artifact removal without suppressing actual image signals.

Main Methods:

• Proposed a new 'aliasing layer' (AL) designed to preprocess MR images by adjusting artifact locations.
• Formulated the AL to work within the spatial locality assumption of CNNs.
• Compared CNNs incorporating ALs against standard CNNs for artifact removal efficacy.

Main Results:

• CNNs utilizing ALs achieved better image-quality metrics compared to deeper CNNs without ALs.
• The AL demonstrated selective suppression of artifacts, preserving genuine image signals.
• A six-layer CNN with ALs outperformed a 12-layer CNN without ALs.

Conclusions:

• The aliasing layer (AL) offers an efficient approach for processing MR-specific artifacts.
• Experimental results confirm that ALs enhance CNN performance in removing artifacts from parallel imaging and EPI.
• This method provides a more effective solution for artifact correction in MR imaging.