Information Extraction from Structural and Functional Brain MR Images using Binary Patterns

本研究的目的在於利用二維特徵法從大腦磁振造影影像中擷取資訊。在現今大數據的時代裡，我們已經有足夠的技術和資源去收集分析世界上所有醫院及研究單位裡的大腦影像；且在多方努力之下，在不久的將來或將成為事實。因此，迫切需要簡單而有效的方法，可以從各種主要的大腦磁振造影影像中擷取資訊，應用於機器學習的演算法上。 本研究利用二維特徵法從三種重要的大腦磁振造影影像中擷取資訊。首先，我們利用區域二維特徵描述大腦磁振造影解剖影像的形態，並將此特徵用來訓練支持向量機模型，使之能夠利用大腦解剖影像分類注意力不足過動症與正常的孩童；結果顯示，利用此法可以達到0.6995的正確率。再者，傳統上比較解剖影像的特徵時，必須先將原始大腦影像轉換至標準化的大腦圖譜上，在相同的圖譜座標系中方能進行比較；在此，區域二維特徵可以直接從未轉換的原始影像上擷取資訊，避免掉一些不必要的轉換和可能引入的雜訊；我們將此法應用在大腦磁振造影解剖影像及擴散磁振造影影像上，並藉此訓練可依大腦結構預測受測者年齡的支持向量機模型；此模型平均絕對誤差最佳可至5.62歲。最後，我們試著把相同的概念套用在功能性磁振造影上，利用自行設計的二維特徵法，藉以描述靜息狀態功能性磁振造影所產生的資料，並據以從正常人中偵測注意力不足過動症與精神分裂症之患者；依此法所學習的模型，其正確率較直接使用傳統分析方法為高，在區分精神分裂症患者與正常控制組上可達0.78，在區分注意力不足過動症患者與正常控制組上可達0.628。 實驗結果顯示，無論是解剖影像、擴散磁振造影影像、或功能性磁振造影影像，皆可使用二維特徵法擷取其中的資訊。由於其簡單且有效的特性，此法相當適合用於未來大規模的大腦科學相關之實驗及研究。This study aimed to build binary methods to extract efficient information from structural brain magnetic resonance (MR) images and functional brain activities. In the era of big data, to collect and analyze all the brain images in hospitals all over the world is technologically possible and might be achieved in the near future. Therefore, simple and effective methods for machine learning algorithms to extract sufficient information from various brain MR images to build classification or regression models based on numerous brain images are critical. In this study, we used binary methods to extract information from three different types of brain MR images. First, we implemented local binary patterns (LBP) to describe anatomical brain morphology and used those patterns to train support vector machine models to classify the attention deficit-hyperactivity disorder (ADHD) subjects from normal ones. As a result, the best accuracy we achieved was 0.6995. Second, different from the traditional methods, which all brain images should be normalized to a standard template to be compared in same atlas coordinates, the LBP was used to extract information from unnormalized brain anatomical images and diffusion tensor imaging. We then constructed age estimation models by that extracted information to show the discriminative power of this approach. The best test result mean absolute error of that model equals 5.62 years. Third, following the same line of thought, a binary mapping method was designed and introduced to detect schizophrenia and ADHD patients using resting-state functional MRI data. Compared with traditional cross-correlation network analysis, proposed models exhibits better performance in detecting schizophrenia and ADHD. Based on our results, the best test accuracy of discriminating schizophrenia from normal subjects was 0.78. The best test accuracy or classifying ADHD from control subjects was 0.628. Results showed those simple binary methods are useful for extract information from structural and functional brain MR images. Those methods are good candidates to be used in large-scale brain science or medicine related researches.口試委員會審定書 i 誌謝 ii 中文摘要 iii ABSTRACT v CONTENTS vii LIST OF FIGURES xv LIST OF TABLES xix Chapter 1 Introduction 1 1.1 Background 1 1.1.1 Brain MR Images 2 1.1.2 Data Driven Method and Machine Learning Methods 5 1.1.3 Brain MR Images and Machine Learning Methods 6 1.1.4 Neuroimaging Data-sharing Initiative 7 1.2 Motivation and Purpose 9 1.3 Voxel Based Morphology (VBM) 11 1.4 Texture Analysis 14 1.5 The structure of the dissertation 14 Chapter 2 Local Binary Patterns (LBP) 16 2.1 The Development and History of the LBP 16 2.2 LBP of 2D Brain MR Images 19 2.2.1 Local Binary Patterns (LBP) 19 2.2.2 Applications in Medical Images and Brain MR Images 22 2.3 Uniform Patterns 23 2.4 LBP of 3D Brain MR Volumes 26 2.4.1 Spatiotemporal LBP - LBP on Three Orthogonal Planes (LBP-TOP) 26 2.4.2 Applications in Medical Images and Brain MR Images 28 2.5 Support Vector Machine (SVM) 28 2.6 Framework of Analysis Using LBP and SVM 29 2.6.1 Build Classification Models Using SVM and LBP-TOP 29 2.6.2 Feature Selection and Building Regression Models Using SVR and LBP 32 Chapter 3 Properties of LBP-TOP 36 3.1 Materials and Methods 37 3.1.1 Participants 37 3.1.2 Parcellations 37 3.1.3 Evaluation 38 3.2 Parameters of LBP 39 3.3 The Efficiency of Uniform Patterns 42 3.4 The Order of LBP Coding 43 3.5 The Effects of Brain MR Image SNR 45 3.6 The Effects of Brain MR Image Resolution 48 Chapter 4 Age Estimation Using Unnormalized MR Brain Images 50 4.1 Introduction 50 4.2 Materials and Methods 53 4.2.1 Subjects and Data Acquisition 53 4.2.2 Data preprocessing 55 4.2.3 Extracting the LBP-TOP Histogram 55 4.2.4 Atlas Registration 58 4.2.5 Support Vector Regression and Feature Ranking 60 4.2.6 Evaluations 62 4.3 Results and Discussion 63 4.3.1 Age estimation using T1WI 63 4.3.2 Age estimation using DTI data 70 4.3.3 Atlas Registration 75 4.3.4 Learned Model 76 4.3.5 Brain Maturation and Aging 77 4.4 Conclusion 80 Chapter 5 ADHD classification using Local Binary Patterns 82 5.1 Introduction 82 5.2 Materials and Methods 85 5.2.1 Participants 85 5.2.2 Diagnostics of ADHD 86 5.2.3 Data preprocessing 89 5.3 Results 99 5.3.1 LBP-TOP 99 5.3.2 Permutation test of basic models 107 5.3.3 Feature selection 108 5.3.4 Resolutions of brain images 110 5.3.5 Tissue types 112 5.4 Discussion 113 5.4.1 Robust to registration method 115 5.4.2 Global effects of ADHD? 116 5.4.3 Combining models using feature selection 117 5.4.4 Most discriminative tissue 117 5.5 Conclusion 118 5.6 Supplements 119 Chapter 6 Discussion, Conclusion, and Future Works 121 6.1 Discussion 121 6.1.1 Structural MRI 121 6.1.2 Different from Traditional Approaches 122 6.1.3 Advantages of Using Binary Patterns and Machine Learning Approaches 124 6.1.4 Knowledge Discovery 124 6.1.5 Limitations 125 6.2 Future Works 127 6.2.1 Normal Ranges as Image Biomarkers of Brain Images 127 6.2.2 Combine Information for Multivariate Approaches 127 6.2.3 Detect ADHD and Schizophrenia Using Functional Connectivity Binary Patterns 128 6.3 Conclusion 165 REFERENCE 16