A Study On Super Resolution Algorithm For Breast Magnetic Resonance Images

Breast cancer is one of the most common woman malignant tumor in the world,and has been the second place in our country with the increase of its incidence and mortality.Early diagnosis and appropriate therapy of breast cancer greatly increased the survival rate of patients.Computer aided diagnosis(CAD)is a kind of new technology which can detect lesions and improve the accuracy of diagnosis by combining computer with medical imaging and digital image processing.Magnetic resonance imaging has been widely used in clinical because of its special advantages,compared with other imaging techniques,such as ultrasound,computed tomography(CT),X-ray.However,the resolution of MRI is limited by hardware condition and scanning time,and these limitations are hard to break through.In the following steps of computer aided diagnosis system,such as image registration,tumor segmentation and classification,highresolution images are necessary.Therefore how to shorten time and recover satisfying resolution with less data is the urgent solution in MRI.At present,a feasible and effective method is introducing super resolution into MRI reconstruction.Super resolution(SR)is an image reconstruction technology proposed in recent years,which uses a group of different fuzzy degree of low-resolution images and additional prior knowledge to reconstruct high-resolution images under the condition of without changing hardware limitations.Recently,domestic and overseas scholars have taken up many SR methods and obtain success on the natural images.However however,application on the monocotyledons plants is still seldom.For Magnetic resonance image(MRI),this paper proposes a single image super-resolution method based on wavelet features and clustered dictionaries.In the training phase,the multiscale wavelet features of low-resolution(LR)images and all high frequency components of high resolution(HR)images are extracted,and then all of these feature images are partitioned into patches.For each class a dictionary is learned using K-Singular Value Decomposition(K-SVD).In the reconstruction phase,each LR patch is classified and sparsely represented with its corresponding dictionary atoms.Iterative Back Projection(IBP)is used for post-processing to further improve the reconstruction quality.Finally the optimal parameters of the whole architecture are given by a serials of contrast experiment with controlled variables.Experimental results show the proposed method outperforms other main-stream methods,both visually and quantitatively.