Intensity non-uniformity correction in MR imaging using residual cycle generative adversarial network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Voxel intensity non-uniformity (INU) is a significant challenge in quantitative magnetic resonance (MR) image analysis, degrading the performance of automated tasks like segmentation and radiomics.
• While visually tolerable, INU can severely impact the accuracy and reliability of quantitative MR image analysis in clinical practice.

Purpose of the Study:

• To develop and evaluate an advanced deep learning-based algorithm for correcting intensity non-uniformity (INU) in magnetic resonance imaging (MRI).
• To enhance the accuracy and efficiency of quantitative MR image analysis by mitigating INU effects.

Main Methods:

• A novel residual cycle generative adversarial network (res-cycle GAN) was developed, integrating residual blocks into a cycle-consistent GAN (cycle-GAN) architecture.
• The model employed an inverse transformation between uncorrected and corrected MRI images for constraint, utilizing a fully convolutional neural network generator for end-to-end transformation.
• The algorithm was trained and evaluated on a cohort of 55 abdominal T1-weighted MR images, comparing its performance against the N4ITK method and other deep learning approaches.

Main Results:

• The res-cycle GAN method demonstrated superior performance with a normalized mean absolute error (NMAE) of 0.011 ± 0.002, peak signal-to-noise ratio (PSNR) of 28.0 ± 1.9 dB, normalized cross-correlation (NCC) of 0.970 ± 0.017, and spatial non-uniformity (SNU) of 0.298 ± 0.085.
• Significant improvements (p < 0.05) in NMAE, PSNR, NCC, and SNU were observed compared to conventional GAN and U-net algorithms.
• The trained model can automatically generate corrected MR images within minutes, eliminating the need for manual parameter tuning.

Conclusions:

• The proposed res-cycle GAN algorithm offers a highly effective and automated solution for correcting intensity non-uniformity in MRI.
• This deep learning approach significantly outperforms existing methods, paving the way for more reliable quantitative MR image analysis in clinical settings.
• The efficiency and accuracy of res-cycle GAN make it a valuable tool for advancing radiomics and other quantitative imaging applications.