| The invention of X-ray CT (computerized tomography) has produced a huge change for clinical diagnosis and treatment. Progress of science and technology has led to the clinical development. Particularly, the medical imaging technology develops rapidly. Currently, the clinical examination and treatment tools have been developed from a single X-ray to a variety of imaging techniques, such as, US (ultrasound), MRI (magnetic resonance imaging), DSA (digital subtraction angiography), DR (digital radiography), PET (positron emission tomography), SPECT (single-photon emission computed tomography), PET-CT, PET-MR, etc. Nowadays, medical imaging technology has become the primary means for the clinical detection and diagnosis of diseases, as well as provides an experimental platform for the clinical research.Medical imaging technology has become increasingly mature. So far, more and more medical imaging devices have been widely used in clinical diagnosis and treatment. Much of the information associated with the patients in clinical medicine can be shown in the medical images, which can assist doctors to correctly diagnose and treatment the diseases. For example, medical images can provide the shape, size, location and other information of the diseased tissues. Specially, CT images have a good display of bone structures and tissue density distributions; US and MR images provide better soft tissue information; PET and SPECT images can reflect the body’s metabolic and function information. Relationship between medical images and clinical diagnosis is becoming closer and closer. Study of medical images has become an important research field which grows rapidly. At present, more and more medical images can be used for clinical diagnosis and research work. Effective analysis (i.e., mining, extraction, integration) of the huge image information is very important for the development of clinical science. Medical image research work mainly includes three aspects:image acquisition, image processing and image analysis. Medical image analysis has an vital role in the medical image research work; it is an important part of medical imaging technology, digital image technology, computer science and technology. Meanwhile, medical image analysis has an important position in medical research, clinical diagnosis and treatment, as well as teaching. With the increasing maturity of medical image analysis techniques, many results of medical image analysis have been used clinically as an important basis for the doctors to diagnose and treat diseases.This study is based on sparse representation and local linear representation to carry out the medical image analysis. In the past 20 years, sparse representation has developed as an important research field in compressed sensing, image denoising, neuroscience, feature extraction, recognition, sample classification, etc. Given an over-complete dictionary, sparse representation aims to use a small number of samples in the dictionary to represent the testing sample, while the representation error is constrained to a small range. In this way, the main information of the testing sample can be extracted, and further used in the processing of the testing sample, such as compression, dimensionality reduction, classification, feature selection, etc. Sparse representation mainly contains three parts:construction of an over-complete dictionary, representation of the testing sample, dictionary sparse coding.1) Construction of the over-complete dictionary. Using training samples to build an over-complete dictionary is the first step of sparse representation. There are two categories of over-complete dictionaries:standard dictionaries and learned dictionaries. Traditional standard dictionaries contains over complete DCT (discrete cosine transform) dictionary, curvelet dictionary, wavelet dictionary, bandlets dictionary, coutourlet dictionary, etc. Currently, learning techniques are usually used in screening huge amounts of training data, thus building compact and expressive learned dictionaries. These algorithms are commonly used in dictionary learning: MOD (method of optimal directions), K-SVD algorithm, method of maximum likelihood,k-means clustering, etc; 2) Representation of the testing sample. LASSO (least absolute shrinkage and selection operator) with L1 norm is the most common form of sparse representation. L1 norm can make most elements of the dictionary coefficients to be zero, and thus ensure the sparsity of the representation. Many algorithms have been derived from LASSO, such as group LASSO, fused LASSO, sparse group LASSO, tree structured group LASSO and overlapping group LASSO; 3) dictionary coding, i.e., solving the dictionary coefficients. Many algorithms can be used to solve the dictionary coefficients, such as, MP (matching pursuit), BP (basis pursuit), OMP (orthogonal matching pursuit), SP (subspace pursuit), ROMP (regularized orthogonal matching pursuit), CoSaMP (compressive sampling matching pursuit), SAMP (sparsity adaptive matching pursuit), StOMP (stagewise OMP), etc.Different from sparse representation, local linear representation does not use a L1 norm to constrain the sparsity of the representation. In local linear representation, a local dictionary is constructed by a small number of closest training samples from the testing sample, and further used to linearly represent the testing sample. There are three steps in local linear representation:1) finding k nearest neighbors of the testing sample from training samples as the dictionary; 2) using the dictionary to linearly represent the testing sample; 3) solving dictionary coefficients. These following algorithms are based on local linear representation:LAE (local anchor embedding), LCC (local coordinate coding), LLE (locally linear embedding), and LLC (locality-constrained linear coding).Three researches were carried out based on sparse representation and local linear representation, as follows:1) Prostate segmentation in 3D CT images. Image-guided radiation therapy for the prostate cancer requires high doses of radiation to the prostate tissue. It is particularly important for accurate segmentation of the prostate volume during the radiotherapy of the prostate cancer. However, distinguishing the prostate from its surrounding tissues is difficult because of the low contrast in CT images. In addition, the position and appearance of the prostate in the CT image of the same patient greatly varies on different treatment days, which brings troubles in accurate segmentation of the prostate in CT images.In this study, an automatic framework is designed for prostate segmentation in CT images. We propose a novel image feature extraction method, namely, VSP (variant scale patch), and design a new classification method called LIP (local independent projection) which is based on local linear representation. Combined with a dictionary online updated strategy, VSP and LIP are used to prostate segmentation in 3D CT images, detailed as follows:Image feature extraction is a key step in medical image analysis. Studies show that the image feature of a point can be effectively represented by its surrounding intensity information. The patch-based image feature of the central point has been widely used in medical image analysis. However, if we want to extract much image information and enlarge the patch size, the dimension of the patch-based feature will increase, and the computing time will increase simultaneously. To overcome the limitation of the patch-based image feature extraction, we propose VSP-based feature extraction algorithm. The main idea is to use the surrounding image information of each point as its feature. Specially, extract accurate image information in the central region and reserve coarse image context in the peripheral area, thus reducing the feature dimension. The proposed VSP feature can provide rich image information in a low dimensional feature space.Based on the basic idea of local linear representation, as well as the independent distribution characteristics of samples from different classes, we propose a LIP-based classification method. In LIP, we assume that samples from different classes lie on different non-linear submanifolds and a a sample can be approximately represented as a linear combination of several nearest neighbors from its corresponding submanifold. Based on this assumption, we use the training samples from each class to represent the testing sample independently, and use LAE, which emphasizes the locality of the representation, to solve the dictionary coefficients. Finally, we label the testing sample to the class with the minimum representation error.Like sparse representation and local linear representation, LIP needs a dictionary which contains training samples. To utilize the latest image information, we use an online updated strategy to construct this dictionary.The proposed method has been evaluated based on 330 prostate 3D CT images of 24 patients. Results show the effectiveness of VSP feature extraction algorithm, LIP classification method, and dictionary online updated strategy in CT prostate segmentations.2) Infant brain MR image registration. MR imaging has high image resolutions, and also can provide good soft tissue information, In addition, the process of MR imaging does not produce X-rays which is harmful to the human body. Using the MR imaging technique, patients do not need to inject contrast agents which may cause allergy, thus using MR imaging in clinics is harmless to patients. Excellent characteristics of the MR imaging technique makes it widely used in hospitals. Accurate registration of infant brain MR images is important for the research on the early brain development and the detection of early brain diseases. However, compared with the adult brain, the infant brain MR images are noisy, also with low image resolution. More seriously, since the infant brain develops rapidly after birth to 1 year old, and the myelin sheath also grows continuously, the intensity difference between the white matter and gray matter in the MR images obtained at two different time points in the first year of life changes a lot, which brings many challenges to the infant brain image registration.This paper proposes an infant brain registration method for MR images obtained at different time points from 0 to 1 year old. The main idea is to utilize growth trajectories learned from a set of longitudinal training images. The learned growth trajectories are able to leverage image registration to tackle the image appearance gap between two infant brain images. Our training dataset contains 24 patients. Each patient has T1 and T2-weighted MR images obtained at five different points between 0 to 1 year old (i.e.,2 weeks,3 months,6 months,9 months,12 months). With the training T1 and T2 MR images obtained at different time points, we apply a 4D segmentation and registration method that integrates complementary multi-modality information to establish the correspondence (i.e., the growth trajectory) between training images at any two different time points.To register two new infant images obtained at different time points, we identify the correspondences between each new image and its respective training images with similar age. Finally, the registration between the two new images can be assisted by the correspondences between each new image with the training images, as well as learned growth trajectories that have been established in the training dataset. To further improve registration accuracy, our infant registration method are combined with a hierarchical and symmetric registration framework that can iteratively refine the registration results.Our dataset contains 24 infant subjects, each with T1 and T2-weighted MR images obtained at 2 weeks,3 months,6 months,9 months,12 months, respectively. Compared to the state-of-the-art methods which can be used in infant brain image registration, the proposed method demonstrated superior registration performance.3) Prediction of CT images from given MR data. Accurate CT predictions from given MR data are clinical desired for dose planning in MR-based radiation therapy and attenuation correction in PET/MR. Methods of the prediction of CT images from given MR images mainly belong to two categories. The first category is atlas registration based methods. These methods highly depend on the accuracy of the deformable registration results. The second category is voxel based methods. Voxel based methods assume a one-to-one correspondence between MR and CT intensities. However, the relationship between MR and CT without any constraint is not a bijection, since materials such as cerebrospinal fluid and air have similar intensities in T1-weighted MR images but different CT values.To solve the common issue in voxel based methods, our study presents a novel method for the prediction of CT images from given MR images based on LSCC (local sparse correspondence combination). In LSCC, we assume that MR samples and CT samples are located on two nonlinear manifolds and the mapping from MR manifold to CT manifold approximates a diffeomorphism under a local constraint. LSCC emphasizes the locality, aiming to ensure the one-to-one correspondence between MR and CT intensities under the local constrains, thus solving the problem in voxel based methods. In LSCC, several techniques are used to strengthen the locality:1) for each point in the testing MR image, a local search window is defined to extract MR training samples (i.e., MR dictionary) from training dataset; 2) κNN (^-nearest neighbors) is used to find k neighbors of the testing point from the MR dictionary; update the MR dictionary to constrain the MR dictionary in a local space; 3) κ-means and κNN are combined to detect outliers in the CT dictionary; delete the outliers to emphasize the locality in CT dictionary; 4) use the MR dictionary to linearly represent the testing point, and use LAE (local anchor embedding), which emphasizes the locality of the representation, to solve the dictionary coefficients. Under these local constraints, the coefficient weights are linearly transferred from MR to CT, and further used to combine the CT dictionary to predict a CT image patch for the testing point. After predicting an image patch for each point in the testing MR image, weight average the overlapped patches to denote the intensity value of each point in the predicted CT image.The proposed method has been evaluated for brain images on a dataset of 13 subjects. Each has T1-, T2-weighted MR images and a CT image with a total of 39 images. In the experiments, mean absolute error of the predicted CT and real CT is calculated to show the effectiveness of the proposed LSCC in the prediction of CT images from given MR images. |
PDF Full Text Request