Kanakaraj P, Yao T, Cai LY, Lee HH, Newlin NR, Kim ME, Gao C, Pechman KR, Archer D, Hohman T, Jefferson A, Beason-Held LL, Resnick SM; Alzheimer’s Disease Neuroimaging Initiative (ADNI); BIOCARD Study Team; Garyfallidis E, Anderson A, Schilling KG, Landman BA, Moyer D. DeepN4: Learning N4ITK Bias Field Correction for T1-weighted Images. Res Sq [Preprint]. 2023 Nov 13:rs.3.rs-3585882. doi: 10.21203/rs.3.rs-3585882/v1. Update in: Neuroinformatics. 2024 Mar 25;: PMID: 38014176; PMCID: PMC10680935.
The study addresses the challenge of removing low frequency intensity artifacts from T1-weighted (T1w) MRI images, a crucial step in ensuring consistent image interpretation. The currently used N4ITK bias field correction, although effective, suffers from integration difficulties across different computational platforms and lacks transparency for optimization. To overcome these limitations, the researchers developed “DeepN4,” a deep learning-based method designed to emulate the N4ITK correction process. This new approach was trained using data from 72 different MRI scanners across various age groups and evaluated against the N4ITK-corrected images, achieving a median peak signal-to-noise ratio (PSNR) of 47.96 dB. Additionally, the DeepN4 model demonstrated strong generalizability across eight external datasets. This innovation not only matches the performance of the traditional method but also enhances flexibility and portability in processing MRI images. All resources related to DeepN4 have been made publicly available, promising broader adoption and further development in the field.
