Research On The Applications Of Sparse Representation In Medical Image Processing

Clinical diagnosis is closely related to human health. Medical images have been widely employed in clinical diagnosis since X-ray was discovered and applied into medical imaging. Nowadays, computer technology and medical imaging techniques are growing rapidly, which brings immense medical data to doctors. Medical image processing can help doctors extract the focal zone from images and give them a better presentation about the information of disease. This technique greatly improves the accuracy and efficiency of clinical diagnosis. The rapid development of clinical methods and strong demands of medical data promote medical image analysis technique to progress continuously.Sparse representation is a hot topic in signal processing. The main idea of sparse representation regards a signal as a linear combination of few atoms in an overcompleted dictionary. With its simple mathematical form and efficient representation ability, sparse representation is widely applied into different research work like image processing and pattern recognition. In the meantime, the application of sparse representation on medical image processing has been concerned by lots of scholars. Sparse representation and sparse regularizer have been applied into many research work of medical image processing.In this thesis, we first made a brief introduction about sparse representation theory. Then the classical approaches in sparse coding and dictionary learning were summarized. Moreover, we applied sparse representation into medical image denoising and breast tumor diagnosis and made detailed discussion about these work.To overcome the impact of noise on image quality in the medical imaging process, images are usually processed by denoising algorithm after acquired from devices. Most of the denoising algorithms utilize the spatial correlation of pixels or frequency transformation to smooth the noisy image, but both of the methods has limited abilities to represent images. In this paper, a sparse representation model is used to model the image. By analyzing the characteristics of noise distribution in medical images, we proposed a patch group based denoising method using SVD. This method not only has a good performance on noise suppression, and can also preserve the local sharpness of images.Breast cancer as a malignant disease for women is hard to cure. The most effective method to control it is to make early diagnosis and prevention. In this paper, we studied the problem about breast tumor images classification. By analyzing the shortcomings of traditional image encoding methods, we improved the dictionary learning methods by incorporating label information into learning process. This method improved the discriminative ability of dictionary without losing its reconstruction ability. In addition, the learned dictionary is applied to represent breast tumor images. Experiments show that our method is robust to different datasets and has a higher classification accuracy compared to the traditional local feature descriptors.