Low-dose Human Brain CT Image Restoration Based On Dictionary Learning In Sparse Gradient Domain

Computed tomography(CT) is widely used in the clinical diagnosis and treatment of diseases. Compared with ordinary X-ray films, human brain CT images have higher density resolution of tissues, and can be used to accurately measure the tiny differences of radiation attenuations between different regions to distinguish the density of each soft tissue. They are highly valued to diagnose craniocerebral diseases. Thus the CT examination is preferred in examining patients with diseases such as cerebrovascular, intracranial tumor, cerebral infarction, and cerebral trauma.The birth and development of multi-slice spiral computed tomography make the imaging quality of CT be improved constantly. However, the X-ray dose is inevitably increased at the same time. Too much radiation dose may result in diseases such as cancer, hence more and more researchers pay attention to low-dose CT images. However, direct reduction of the X-ray dose in CT would lead to a quality loss of the images and have an impact on the doctor’s diagnosis. On the premise of guaranteeing the quality of CT images, how to effectively reduce the X-ray radiation dose has become an important topic and a research focus in the field of medical images.Owing to its excellent performance, the sparse representation method based on dictionary learning has been applied to signal processing problems such as image denoising and image restoration. Some researches show that the improvement of the sparsity of the training samples can improve the efficiency and robustness of the learned dictionary to a certain extent. Compared with the original images, the corresponding gradient images are sparser because there exists a strong correlation between adjacent pixels. So dictionary learning algorithm based on sparse gradient domain would improve the learning efficiency of the dictionary. Unfortunately, the introduction of gradient operator may enlarge the noise in most cases.In order to reduce the influence of noise amplification caused by the gradient operator, this work presents two improved low-dose human brain CT image restoration algorithms based on dictionary learning in sparse gradient domain. In one of the algorithms, the principal component analysis(PCA) is carried out on a gradient image first to alleviate the bad effects of the gradient operator and then the PCA transformed data are used to train dictionary in order to reach denoising. Another algorithm filters the CT image blocks via three-dimensional block matching(BM3D) before the gradient operations. The gradient images of the filtered CT are applied to learning dictionary. Experiments on the clinical human brain CT images show that both algorithms have desired denoising performance on low-dose CT images, which is expected, if applied in clinics, to be helpful to greatly reduce the X-ray radiation dose received by the patients without affecting the doctors’ diagnosis accuracy.