Development and evaluation of a deep learning model for multi-frequency Gibbs artifact elimination

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Gibbs artifacts, caused by k-space truncation, degrade MRI image quality.
• These artifacts can be mistaken for syrinx, complicating diagnosis.
• Current methods for artifact removal are insufficient.

Purpose of the Study:

• Develop and validate a deep learning (DL) model for multi-frequency Gibbs artifact elimination.
• Assess the model's performance across diverse anatomical regions, MRI sequences, and artifact severities.
• Evaluate the impact of artifact removal on syrinx diagnosis.

Main Methods:

• Retrospective analysis of 290,940 MRI images from 4,936 scans.
• Training the DL model with artificially generated Gibbs artifacts.
• Prospective validation using data from 20 healthy adults and 10 syrinx patients.

Main Results:

• Processed images showed significantly higher quality than original or conventionally filtered images (P<0.05).
• Syrinx identification confidence increased with DL model use (AUC: 0.95 vs. 0.90, P=0.04).
• The model demonstrated robust Gibbs artifact removal across various conditions.

Conclusions:

• The DL model effectively eliminates Gibbs artifacts.
• It shows potential for enhancing syrinx identification accuracy.
• The model is robust and performs well in real-world clinical scenarios.