基于磁共振圖像的阿爾茨海默病神經(jīng)指紋研究
發(fā)布時(shí)間：2018-07-15 15:56
【摘要】：隨著全球逐步步入老齡化,阿爾茨海默病(Alzheimer Disease,AD)漸已成為當(dāng)今社會(huì)威脅人類身體健康、家庭生活質(zhì)量和社會(huì)良性發(fā)展的最為嚴(yán)重的疾病之一。但是鑒于目前阿爾茨海默病只能延緩無法做到完全治愈,因此對于阿爾茨海默病的早期診斷也就愈發(fā)重要。目前對于AD的診斷仍然依賴于臨床癥狀和專業(yè)醫(yī)師的主觀判斷,磁共振成像技術(shù)(Magnetic Resonance Imaging,MRI)的出現(xiàn)提供了一種非介入式的無創(chuàng)腦疾病檢查方式。雖然目前利用MRI觀察AD患者發(fā)現(xiàn)了諸如海馬體積變化等現(xiàn)象,但由于多腦區(qū)圖像的分割精度過低,因此尚未完全解決AD的特征量化問題。本研究基于多腦圖譜自動(dòng)分割,從圖譜預(yù)選擇方法出發(fā),完成對彌散張量圖像的分割并研究感興趣區(qū)域的紋理特征,利用所選腦結(jié)構(gòu)和紋理特征構(gòu)建AD神經(jīng)指紋模型,為AD的研究提供新的思路。由于AD在影像學(xué)上沒有顯著的病灶區(qū),對AD的MR圖像量化分析時(shí)就需要針對大腦的各組織結(jié)構(gòu)進(jìn)行具體研究,因此,對于大腦組織結(jié)構(gòu)進(jìn)行準(zhǔn)確的分割也就顯得尤為重要。針對目前分割精度最高的基于多圖譜的圖像分割結(jié)果精度尚有提升空間的現(xiàn)狀,本文在多圖譜圖像分割方法的圖譜選擇階段,提出了兩種新的圖譜預(yù)選擇方法,一種是利用側(cè)腦室結(jié)構(gòu)標(biāo)簽進(jìn)行圖譜預(yù)選擇,另一種是將側(cè)腦室、腦白質(zhì)、腦灰質(zhì)以及腦脊液四個(gè)結(jié)構(gòu)標(biāo)簽融合為一個(gè)新標(biāo)簽圖譜進(jìn)行圖譜預(yù)選擇,并使用約翰霍普金斯大學(xué)影像中心的圖譜數(shù)據(jù)庫和圖像自動(dòng)分割方法,實(shí)現(xiàn)了對于T1腦圖像的分割,實(shí)驗(yàn)結(jié)果表明,本文提出的兩種圖譜預(yù)選方法提高了圖像的分割精度的同時(shí)縮短了分割時(shí)間,為多模態(tài)參數(shù)圖的分割并構(gòu)建神經(jīng)指紋奠定了基礎(chǔ)。在利用新的圖譜預(yù)選擇方法實(shí)現(xiàn)了對T1圖像精準(zhǔn)分割的基礎(chǔ)上,將T1結(jié)構(gòu)圖像分割結(jié)果映射到彌散張量成像(Diffusion Resonance Imaging,DTI)多模態(tài)參數(shù)圖上,實(shí)現(xiàn)了對多模態(tài)參數(shù)圖的精準(zhǔn)分割。針對目前已有AD的量化特征不明顯的現(xiàn)狀,本文選取AD患者腦結(jié)構(gòu)性狀發(fā)生改變的28個(gè)重點(diǎn)區(qū)域,提取了13個(gè)紋理特征并利用過濾法進(jìn)行特征篩選,最終針對三類多模態(tài)參數(shù)圖確定了10個(gè)特征和23個(gè)感興趣區(qū)域,構(gòu)建了具有顯著意義的AD神經(jīng)指紋模型。本文的研究成果表明了通過改進(jìn)圖譜預(yù)選擇方法可以有效提高多圖譜磁共振圖像的分割精度,同時(shí)驗(yàn)證了結(jié)合多模態(tài)參數(shù)圖提取紋理特征構(gòu)建神經(jīng)指紋的可行性,為針對AD進(jìn)行更全面的研究奠定了科學(xué)基礎(chǔ)。
[Abstract]:With the aging of the world, Alzheimer disease (AD) has become one of the most serious diseases threatening human health, family life quality and social benign development. But given that Alzheimer's can only delay a complete cure, early diagnosis of Alzheimer's is becoming increasingly important. At present, the diagnosis of AD still depends on clinical symptoms and subjective judgment of professional doctors. Magnetic Resonance Imaging (MRI) provides a non-interventional non-invasive examination of brain diseases. Although MRI has been used to observe AD patients, such as hippocampal volume changes, but the segmentation accuracy of multi-brain images is too low, so the problem of AD feature quantization has not been completely solved. Based on the automatic segmentation of multi-brain atlas, the segmentation of diffuse Zhang Liang images and the study of texture features of regions of interest were completed based on the pre-selection method, and the AD neural fingerprint model was constructed by using the selected brain structure and texture features. To provide a new idea for the study of AD. Since AD has no obvious focus area on imaging, it is necessary to study the structure of brain tissue in the quantitative analysis of AD image. Therefore, it is very important to segment the structure of brain tissue accurately. In view of the fact that there is still room for improvement in the accuracy of image segmentation based on multi-atlas which has the highest segmentation accuracy at present, this paper proposes two new pre-selection methods for multi-atlas image segmentation in the phase of spectrum selection of multi-atlas image segmentation method. One is to preselect the map by using the label of the lateral ventricle structure, the other is to fuse the four structural labels of lateral ventricle, white matter, gray matter and cerebrospinal fluid into a new label for pre-selection. Using the map database of Johns Hopkins University Image Center and the automatic image segmentation method, the T1 brain image segmentation is realized. The experimental results show that, The two methods proposed in this paper not only improve the accuracy of image segmentation but also shorten the segmentation time, which lays a foundation for the segmentation of multimodal parameter images and the construction of neural fingerprints. Based on the accurate segmentation of T1 images by using a new method of map pre-selection, the segmentation results of T1 structure images are mapped to the multimodal parameter diagrams of Diffusion Resonance Imaging (Zhang Liang), and the precise segmentation of multimodal parametric images is realized. In view of the fact that the quantitative characteristics of AD are not obvious at present, this paper selects 28 key regions in which the brain structural traits of AD patients change, and extracts 13 texture features and selects them by filter method. Finally, 10 features and 23 regions of interest are determined for the three types of multimodal parameter maps, and a significant AD neural fingerprint model is constructed. The research results of this paper show that the segmentation accuracy of multispectral magnetic resonance image can be improved effectively by improving the pre-selection method of map. At the same time, the feasibility of constructing neural fingerprint by extracting texture feature from multi-modal parameter graph is verified. It lays a scientific foundation for a more comprehensive study of AD.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2016
【分類號】：R749.16;R445.2;TP391.41
[Abstract]:With the aging of the world, Alzheimer disease (AD) has become one of the most serious diseases threatening human health, family life quality and social benign development. But given that Alzheimer's can only delay a complete cure, early diagnosis of Alzheimer's is becoming increasingly important. At present, the diagnosis of AD still depends on clinical symptoms and subjective judgment of professional doctors. Magnetic Resonance Imaging (MRI) provides a non-interventional non-invasive examination of brain diseases. Although MRI has been used to observe AD patients, such as hippocampal volume changes, but the segmentation accuracy of multi-brain images is too low, so the problem of AD feature quantization has not been completely solved. Based on the automatic segmentation of multi-brain atlas, the segmentation of diffuse Zhang Liang images and the study of texture features of regions of interest were completed based on the pre-selection method, and the AD neural fingerprint model was constructed by using the selected brain structure and texture features. To provide a new idea for the study of AD. Since AD has no obvious focus area on imaging, it is necessary to study the structure of brain tissue in the quantitative analysis of AD image. Therefore, it is very important to segment the structure of brain tissue accurately. In view of the fact that there is still room for improvement in the accuracy of image segmentation based on multi-atlas which has the highest segmentation accuracy at present, this paper proposes two new pre-selection methods for multi-atlas image segmentation in the phase of spectrum selection of multi-atlas image segmentation method. One is to preselect the map by using the label of the lateral ventricle structure, the other is to fuse the four structural labels of lateral ventricle, white matter, gray matter and cerebrospinal fluid into a new label for pre-selection. Using the map database of Johns Hopkins University Image Center and the automatic image segmentation method, the T1 brain image segmentation is realized. The experimental results show that, The two methods proposed in this paper not only improve the accuracy of image segmentation but also shorten the segmentation time, which lays a foundation for the segmentation of multimodal parameter images and the construction of neural fingerprints. Based on the accurate segmentation of T1 images by using a new method of map pre-selection, the segmentation results of T1 structure images are mapped to the multimodal parameter diagrams of Diffusion Resonance Imaging (Zhang Liang), and the precise segmentation of multimodal parametric images is realized. In view of the fact that the quantitative characteristics of AD are not obvious at present, this paper selects 28 key regions in which the brain structural traits of AD patients change, and extracts 13 texture features and selects them by filter method. Finally, 10 features and 23 regions of interest are determined for the three types of multimodal parameter maps, and a significant AD neural fingerprint model is constructed. The research results of this paper show that the segmentation accuracy of multispectral magnetic resonance image can be improved effectively by improving the pre-selection method of map. At the same time, the feasibility of constructing neural fingerprint by extracting texture feature from multi-modal parameter graph is verified. It lays a scientific foundation for a more comprehensive study of AD.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2016
【分類號】：R749.16;R445.2;TP391.41
【參考文獻(xiàn)】
相關(guān)期刊論文 前7條
1 湯惠;王麗萍;魯軍;潘云龍;;國內(nèi)阿爾茨海默病樣本收集現(xiàn)狀分析[J];中國醫(yī)藥生物技術(shù);2015年06期
2 江貴平;秦文健;周壽軍;王昌淼;;醫(yī)學(xué)圖像分割及其發(fā)展現(xiàn)狀[J];計(jì)算機(jī)學(xué)報(bào);2015年06期
3 夏，
本文編號：2124604
本文鏈接：http://sikaile.net/kejilunwen/ruanjiangongchenglunwen/2124604.html
最近更新
教材專著
