Research On Magnetic Resonance Image Denoising Algorithm

Magnetic resonance(MR)images reveal the internal structure of the human body.It can assist doctors in finding the location of lesions.However,due to the limitation of Magnetic Resonance Imaging(MRI)technology,MR images contain a lot of Rician noise that affects the quality of images.In severe cases,doctor misdiagnoses the condition because images’ lesions are blurred.Medical images need the help of detailed information such as tissue structure contours when segmenting the target.Noise makes the detailed information lost,resulting in a decrease in segmentation accuracy.In order to solve the problems of noisy MR images and limited improvement in segmentation accuracy,this paper proposes two denoising algorithms for the Rician noise contained in MR images.The main research contents are as follows:(1)Research on MR images denoising algorithm combining variance stable transform and PGPDThe noise contained in MR images obeys the Rician distribution.There is a strong dependence between noise and signals.As a representative of the traditional denoising algorithm,PGPD(Patch Group Based Prior Learning Denoising)algorithm can well remove Gaussian noise.However,it fails to achieve the desired result when the PGPD algorithm is used to remove the Rician noise from MR images.Therefore,a new denoising algorithm VST-PGPD(Variance Stable Transformation-Patch Group Based Prior Learning Denoising)is proposed.Rician noise is transformed into noise approximately obeying Gaussian distribution by variance stable transformation.PGPD algorithm is used to denoise in the transform domain.Through experiments on MR images denoising of different slices and different tissues,comparing with the PGPD algorithm,the VST-PGPD algorithm can effectively remove the Rican noise contained in MR images,and obtain higher peak signal-to-noise ratio(PSNR)and similar structure(SSIM).(2)Research on MR images denoising algorithm based on FFDnet networkFor digital image denoising,deep learning algorithms perform better than traditional algorithms.Most of the research objects are Gaussian noise,but there is less research on Rician noise removal in MR images.Based on FFDnet(Fast and Flexible Solution for CNN)network,a new denoising algorithm MFFD(MR image denoising based on FFDnet)is proposed.By using the actual MR image from the IXI dataset and modifying the noise type,MFFD is used to train denoising models with different Rician noise levels.Through setting the number of different network layers to find the best structure for Rician noise removal in MR images and comparing the denoising results of BM3 D,RED-WGAN and Mc Dn CNN,MFFD is the best of them.Experiments are performed on MR images with mixed noise and actual MR images to verify the effectiveness and robustness of the MFFD algorithm.(3)Effect of noise in MR images on prostate target segmentationBecause the intensity and type of noise contained in the actual MR image is unknown,it is difficult to evaluate the degree of influence on image target segmentation.An experiment of prostate target segmentation in MR images is introduced.We obtain data from the publicly available data set of the PROMISE12 Challenge.U-Net network is used to segment prostate.Dice coefficient(DSC)and recall are used as evaluation indicators.There are two research purposes for experiments The one is to study the effect of actual noise on prostate target segmentation based on the segmentation results of the actual MR images and the MFFD algorithm to denoise the image,the other is to study the effect of superimposed noise on prostate target segmentation by comparing MR images of different levels of Rican noise with MFFD algorithm before and after denoise segmentation results.Through the above experiments,it is verified that the MFFD algorithm can effectively remove the noise contained in the MR image and improve the accuracy of target segmentation.