DeepN4: Learning N4ITK Bias Field Correction for T1-weighted Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• T1-weighted (T1w) MRI images suffer from low-frequency intensity artifacts caused by magnetic field inhomogeneities.
• Bias field correction is crucial for consistent interpretation of T1w MRI data.
• Current state-of-the-art N4ITK bias correction lacks portability and flexibility, hindering reproducibility.

Purpose of the Study:

• To develop a portable, flexible, and fully differentiable deep learning approximation of N4ITK bias field correction.
• To create a method that facilitates methodological development and improves cross-platform reproducibility.

Main Methods:

• A deep learning network, termed DeepN4, was trained using N4ITK-corrected T1w MRI and bias fields from eight independent cohorts (72 scanners).
• Supervision was performed in log space.
• The network's performance was evaluated using peak signal-to-noise ratio (PSNR) and tested on external datasets.

Main Results:

• DeepN4 closely approximated N4ITK bias field correction.
• The median PSNR between N4ITK and DeepN4 corrected images was 47.96 dB.
• The DeepN4 model demonstrated generalizability across eight external datasets.

Conclusions:

• Naïve deep neural networks can effectively approximate N4ITK bias field correction.
• DeepN4 offers a flexible and portable alternative to N4ITK, enhancing reproducibility in MRI preprocessing.
• The developed method and code are publicly available to facilitate further research.