Residue Learning Deep Discriminative Network for 3D Magnetic Resonance Image Denoising and Bias-field Correction

Abstract

Random noise and bias field artifacts in magnetic resonance images (MRI) significantly degrade im-age quality and destroy essential spatial image information. These factors result in uniform distortions, smoothing variations in low-frequency MR signals, and irregular intensity distributions, severely affect-ing the original image data across the image regions. The magnetic bias field effects introduce varying intensity variations among tissue regions with similar physical properties that cause dark shading effects, specifically in low-resolution image regions. MR images affected with noise and bias-field or intensity in-homogeneity result in inaccurate interpretation, flawed analysis, incorrect segmentation, and misinformed visual understanding. This paper introduces a deep learning (DL) based feature-aware residue-learning encoder-decoder generator convolutional neural network (RLDCNN) model with a discriminator module to restore quality images from real and fake generated images. The generator model reproduces image information in high-frequency intensity regions using auto-encoding operations. The experimental per-formance of the proposed method shows minimization of information loss, improved peak-signal-to-noise ratio (PSNR), and structural similarity index measurement (SSIM) values on publicly available brain MRI datasets, BrainWeb, and IXI. The experimental results significantly show improved performance compared to existing MRI denoising methods with 37.83 and 36.35 averaged PSNR, 0.962 and 0.965 averaged SSIM with Brainweb and IXI real MRI dataset with noise levels from 3% to 15% respectively. The averaged bias-field correction result achieved at 34.00 PSNR and 0.974 SSIM in bias corrected images which is significantly higher than the existing methods.