Research On Image Classification And Extraction For Traumatic Brain Injury Based On Sparse Representation Model

Visualization of traumatic brain injury (TBI) can improve the accuracy and efficiency of the diagnosis of many brain diseases. However, computer aided diagnosis of brain injury is also facing many challenges, such as the complex structure of the brain and the low gray contrast of the tissues in the MR images, etc.The computer-aided diagnosis system of TBI is mainly related to image classification and object detection and extraction. Image classification and object detection and extraction is currently an important research direction in the field of pattern recognition, and has important applications in human-computer interaction, computational intelligence and other fields. A theory of sparse representation was proposed according to the sparse characteristic of brain images, and this becomes a new signal representation and has been widely used in signal processing, computer vision analysis and other fields. When a signal is represented in sparse representation, it is decomposed into a linear combination of over-complete dictionary, in which the coefficient vector is sparse. Sparse representation is more robust against signal noise. This thesis focuses on sparse representation, and also focuses on classification and object detection of MR images. Based on related references and existing research, further in-depth study was carried out. In this thesis, the work can be summarized as follows:First of all, I summarize the basics of traumatic brain injury and the characteristics of magnetic resonance images, followed by detailing the importance of the research on TBI at this stage and the difficulties in processing images with traumatic brain injury. Then I introduce the domestic and international research status of the image classification and object detection, and summarize a detailed analysis of the mathematical model based on sparse representation. This thesis describes the convex optimization algorithm and greedy algorithm, and then introduces the classification of MR images based on sparse representation. Secondly, this thesis studies the classification methods of MR images to determine whether an object is with TBI. The classification methods of high-dimensional pattern, e.g., support vector machines (SVM), have been widely investigated for analysis of structure and function of brain images (such as MRI) to assist the diagnosis of TBI. Most existing classification methods extract features from MR data and then construct a single classifier to perform classification. However, due to noise and small sample size of MR data, it is a challenge to train a single global classifier that can be robust enough to achieve a good classification performance. In this thesis, we investigate the feasibility of using MRI-based textures to classify subjects with and without TBI. We propose a local patch-based subspace ensemble method which builds multiple individual classifiers based on different subsets of local patches and then combines them for more accurate and robust classification. We divide each brain image into a number of local patches and select a subset of patches randomly from the patch pool to build a weak classifier. Here, the sparse representation-based classifier (SRC) method is used to construct each weak classifier. Then, we combine multiple weak classifiers to make the final decision. The results show that the local patch-based subspace ensemble classification using MRI textures can effectively classify subjects/slices with and without TBI.Thirdly, this thesis studies how to use RPCA algorithm to segment the injured area. The existing extraction methods of the TBI area, still based on the traditional segmentation methods. However, there is not yet an effective segmentation method that can achieve a good extraction performance of injured area because each method has its own limitation and MR images have specific properties. In order to achieve automatic, rapid and accurate detection and extraction of injured area, this thesis presents a robust principal component analysis algorithm (RPCA) based on sparse representation. The final results show that this algorithm is more robust and faster than the traditional algorithms.