Deep convolutional neural networks for bias field correction of brain magnetic resonance images

As a low-frequency and smooth signal, the bias field has certain destructive effect on magnetic resonance (MR) images is main obstacle for doctors' diagnosis image processing (such as segmentation, texture analysis, registration). Before analyzing damaged MR image, preprocessing step required to correct in image. Unlike traditional removal algorithms based signal models priori assumptions, deep learning methods do not require precisely modeling signals fields need adjust parameters. An with input corrected output after neural network being trained large training set. In this paper, we propose taking original local feature of multiple frequency bands obtained by Log-Gabor filter bank input, correcting brain through separable convolutional network. Meanwhile, speed up process improve correction performance, apply residual batch normalization. We conducted same test BrainWeb simulation database Human Connectome Project real data set, consistency qualitative quantitative evaluation shows that our model demonstrates better performance than state-of-the-art N4 non-iterative multi-scale (NIMS) methods. Especially high-intensity non-uniformity level, been well corrected.