並列タイトル等非アルコール性脂肪肝疾患における機械学習技術を用いたRaman組織学
一般注記Histopathology is a standard means to diagnose the disease states of cells or tissues fromtheir morphological features, but it requires the expertise of histopathologists and is,therefore, susceptible to human bias. Raman micro-spectroscopy can provide additionalbiochemical information that is not available by morphological examination, and hasa large potential to assist histological inspection and to benefit diagnosis of disease asobjectively as possible in fluorescent label-free manner. Detailed analysis of Raman mi-croscopic data is essential to detect the spectral changes originating from the underlyingbiochemical changes in cells or tissues due to the progression of disease.This thesis is concerned with the development of diagnostic tools by integrating Ra-man microscopic imaging with methods of machine learning and information theory,and the analysis of Raman hyper-spectral images of rat liver tissues comprising a di-etary model of non-alcoholic fatty liver disease (NAFLD), with each liver tissue havingbeen histopathologically diagnosed as normal, non-alcoholic fatty liver (NAFL), or non-alcoholic steatohepatitis (NASH).In the first study, dimension reduction (manifold learning) and ensemble-learning-based random forest classification are performed on the Raman spectra obtained fromthe regular spatial grid averaging of Raman images for predicting different states (dietaryand histological). I identify a set of important Raman bands in differentiating the Ramanspectra arising from different states of tissues. Furthermore, I find that NAFL state isdistinguished into two pahses, namely, ‘slowly progressive NAFL’ (NAFL-α) and ‘rapidlyprogressive NAFL’ (NAFL-β) in terms of Raman imaging, and main Raman shifts toseparate these two NAFL models are identified. This enhances the diagnostic capabilitiesto distinguish the states of tissues at early stages of the disease.In the second study, using the dietary model of NAFLD in rats, I apply machine learn-ing and information theory to evaluate cellular-level information in liver tissue samples. The method first increases the signal-to-noise ratio while maintaining spatial and spectralstructures of Raman images as much as possible through extension of the simple lineariterative clustering superpixel algorithm developed in the area of image analysis. Second,using the unsupervised machine learning with rate distortion theory and the Poisson errorarising from photon counting, it identifies a set of characteristic spectra having distinctRaman information across the tissues. I discover diverse chemical environments in theliver tissues, allowing for the quantification of representative biochemical componentssuch as vitamin A, lipids, and cholesterols which can be very important insights intothe disease states of cells or tissues. Third, armed with microscopic information aboutthe biochemical composition of the liver tissues, I group tissues having similar chemicalcomposition using agglomerative hierarchical clustering, providing a novel “descriptor”enabling us to infer tissue states, contributing valuable information to histological inspec-tion. Excessive lipid deposition with the appearance of cholesterol signatures indicatesthe severity of the disease state of the NAFLD tissues.Raman microscopy coupled with the proposed techniques will offer new clinical toolsthat will aid pathologists in more precise NAFLD diagnosis with molecular informationabout the liver tissue, and enable us to predict the progression of disease at some earlystages where any morphological feature of diseases does not appear yet.Key words:Raman Hyper-Spectral Imaging, Non-Alcoholic Fatty Liver Disease, ManifoldLearning, Machine Learning, Superpixel Segmentation, Rate-Distortion Theory.
(主査) 教授 小松崎 民樹（電子科学研究所）, 教授 芳賀 永, 教授 出村 誠, 准教授 原田 義規（京都府立医科大学大学院医学研究科）
生命科学院（生命科学専攻）
コレクション（個別）国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
受理日（W3CDTF）2020-02-10T02:20:18+09:00
連携機関・データベース国立国会図書館 : 国立国会図書館デジタルコレクション