Feature extraction for tau images for evaluating Alzheimer’s disease with texture analysis

• Main Authors: Song-TingLee, 李松庭
• Other Authors: Yu-Hua Fang
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/3cwjxf

碩士 === 國立成功大學 === 醫學工程研究所 === 105 === Alzheimer’s disease (AD) is the most common cause of dementia with no approved treatment or cure up to date. Development of drug strategies focuses on the abnormal production and clearance of the amyloid-β(Aβ) peptides and tau proteins. So far, all AD drugs underwent Phase III trial have failed to show significant efficacy. There are many potential reasons for such failure of clinical trials. One of the reasons, and perhaps the most important one, is the lack of identification for the pathological status of a test subject. There is an urgent need to find accurate methods of early detection and effective therapies for AD before the symptoms start. Recently, noninvasive detection of tau proteins using positron emission tomography (PET) with the advent of a tau tracer 18F-AV1451 is aimed to assist the diagnosis of AD as well as to track and predict the disease progression. Quantitative analysis of tau pathology in human brain with tau images can be a powerful method as a diagnostic aid for staging the disease. Therapies, especially those targeting irreversible neurodegenerative processes, may have a better chance of succeeding if applied early and specifically. Methods: This study is focused on 18F-AV1451 PET images quantitative analysis. We acquired 141 tau image data from 127 patients with clinical diagnosis information from Alzheimer’s Disease Neuroimaging Initiative (ADNI) and Department of Defense ADNI (ADNIDOD). We have developed a fully automatic quantitative method based on SPM8, and used Chang Gung Image Texture Analysis (CGITA) toolbox to evaluate the tau images with texture analysis. Furthermore, we have added additional wavelet-based methods into CGITA for generating the quantitative textural features. With our methods, a large amount of quantitative textural features can be extracted after the brain segmentation. We evaluated the relation between the cognitive functions and textural features by area under curve (AUC) of receiver operation characteristics (ROC) curves. Results: In the ROC analysis, we found that, when using the mini-mental state examination (MMSE) cutoff of 23 to discriminate cognitive normal (CN) from cognitive impairment (CI), the AUC is 0.851 for wavelet-based entropy, with 71.4% sensitivity and 89.1% specificity. The classification of normal (NL) and AD groups yields an AUC of 0.82 for standard uptake value (SUV) Entropy-Asphericity product, with 81.8% sensitivity and 76.56% specificity. Conclusion: The wavelet-based features may have a great potential to be used for early detection of CI patients. With our fully automatic quantification methods, we may further extend this application into oncology and cardiology for image-based discrimination and classification.