Rician Distribution Neural Network Based On Variance-Stabilization Transformation

Magnetic resonance(MR)images are often corrupted by noise during the imaging process,and such degraded images can affect subsequent image processing and cause medical misdiagnosis.As a result,denoising techniques for MR images are of great importance.We usually use the Rician distribution to model MRI noise images.The Variance-Stabilization Transformation(VST)method of the Rician distribution image is performed by converting the noise signal into a variance-stabilized signal so that its noise can be considered as additive noise,then the conventional denoising algorithm designed for AWGN can be used upon the transformed signal.The final clean image is obtained by using an Exact Unbiased Inverse(EUI)transformation.In this paper,a deep learning neural network method based on variance stabiliza-tion theory is proposed for solving the denoising problem of Rician distribution.By studying the workflow of the traditional VST method and the construction of the EUI transform,this paper proposes a CNN method fitting the variance-stabilized transfor-mation,which is combined with a Gaussian denoiser obtained by pre-training natural high definition images.Our method achieves a Rician distribution denoiser that out-performs the traditional VST method.In addition,we propose an end-to-end network structure based on the VST workflow and optimize the network learning strategy by using Local Standard Deviation(LSD)feature maps to obtain better denoising results.The experimental study under MRI brain image dataset verifies the effectiveness of our approach in controlling the image distribution variation and solving the Rician distribu-tion denoising problem.