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

Gibbs artifacts frequently occur as a result of truncation in the frequency domain (k-space). Gibbs artifacts can degrade image quality and may be misinterpreted as syrinx, thereby complicating the diagnosis. This study aimed to develop and evaluate a robust deep learning (DL) model that eliminates multi-frequency Gibbs artifacts.

We retrospectively collected 290,940 magnetic resonance imaging (MRI) images from 4,936 scans, encompassing 5 anatomical regions and 67 MRI sequences, to develop a DL model for Gibbs artifact removal. This model was trained using artificially generated Gibbs artifacts, featuring various truncation ratios as input data, and its performance in artifact removal was evaluated across different anatomical regions, MRI sequences, and levels of Gibbs artifact severity. For external validation, we prospectively collected data from 20 healthy adults and 10 syrinx patients, comparing radiologists' diagnostic accuracy with area under the receiver operating characteristic curves (AUC) on images before and after artifact removal to assess the model's impact on syrinx diagnosis.

The images processed by our model demonstrated a statistically significantly higher image quality score than the original images and those processed by conventional filtering algorithms (all P<0.05). Moreover, the model enables greater confidence in identifying syrinx compared to the original images [AUC: 0.95, 95% confidence interval (CI): 0.92-0.99] versus 0.90 (95% CI: 0.86-0.95) (P=0.04).

The model demonstrates excellent performance and robustness in eliminating Gibbs artifacts and may hold the potential for improving syrinx identification.