進階搜尋


   電子論文尚未授權公開,紙本請查館藏目錄
(※如查詢不到或館藏狀況顯示「閉架不公開」,表示該本論文不在書庫,無法取用。)
系統識別號 U0026-2501202118403000
論文名稱(中文) 以T1權重海馬迴磁振影像與三維卷積神經網路分類阿茲海默症
論文名稱(英文) Classification of Alzheimer’s disease in T1-weighted Hippocampal Images with 3D Convolutional Neural Network
校院名稱 成功大學
系所名稱(中) 資訊工程學系
系所名稱(英) Institute of Computer Science and Information Engineering
學年度 109
學期 1
出版年 110
研究生(中文) 林紹弘
研究生(英文) Shau-Hung Lin
學號 P76064033
學位類別 碩士
語文別 英文
論文頁數 47頁
口試委員 指導教授-吳明龍
口試委員-莊子肇
口試委員-趙梓程
中文關鍵字 三維卷積神經網路  ADNI  阿茲海默症  海馬迴  海馬迴子區域 
英文關鍵字 3D convolutional neural network  ADNI  Alzheimer’s disease  hippocampus  hippocampal subfields 
學科別分類
中文摘要 目的:阿茲海默症(AD)是一種漸進性的神經退化性疾病,在阿茲海默症患者的腦中,海馬迴組織的體積會大幅縮小。本研究採用ADNI資料集的T1權重海馬迴影像與三維卷積神經網路(3D CNN)來分類阿茲海默症患者與一般人。此資料集包含420個來自不同受測者的影像,其中AD與CN的數量相等。由於卷積神經網路使用影像上的特徵來進行分類,本研究亦試圖理解哪些影像特徵會對分類阿茲海默症造成影響。

方法:影像先以FreeSurfer軟體進行前處理,照順序包含解析度調整、亮度的正規化、計算對MNI305腦地圖的曲射變換及分割不同結構。接著海馬迴部分的影像會被抽取出來。這些影像透過旋轉來進行資料增強。而一個支援向量機(SVM)分類器被作為比較準確度的基準。本研究採用的神經網路包含四個卷積區塊,左側和右側的網路在數個卷積層後序連以得出結果。神經網路以六種不同的模型以及兩種不同的輸入進行訓練。除了第一個模型外,海馬迴體積與受試者的年齡皆被作為額外特徵輸入神經網路中,而各個模型採用不同的方法來將之與神經網路結合。第一種輸入影像僅包含海馬迴的影響,而第二種輸入則是一個包含海馬迴周邊區域及海馬迴遮罩的雙通道影像。訓練時採用K-fold交叉驗證,其中K=5。在完成卷積神經網路的訓練後,會進行特徵圖的反推。首先是從神經網路後段的序連層中選出重要項,反推出各個卷積層中特徵圖的貢獻度,從而獲得輸入影像各個體素的貢獻度。最後,此一貢獻度與海馬迴的子區域進行比較,以了解每個子區域對於分類上的影響力。

結果:不同模型的驗證資料準確度介於0.79至0.85之間,最高者為0.85,其標準差0.03。該模型將透過支援向量機訓練過的額外參數與神經網路結果相加,並且使用僅有海馬迴的輸入影像。而測試資料的準確度則介於0.78至0.85之間,最高者為0.85,其標準差為0.03。該模型將額外資料與神經網路輸出序連,並使用包含海馬迴周邊區域及海馬迴遮罩的雙通道影像。反推神經網路的結果顯示出海馬迴的CA1與subiculum子區域對AD分類上有著顯著的影響。

結論:比較各個模型後,結果顯示出將海馬迴的體積與受試者的年齡輸入神經網路並不能顯著的改善分類結果。而有著高初始準確度的模型則有著的過擬合問題。而反推特徵的結果指出CA1與subiculum區域對AD分類有著顯著的影響,此一結果也符合其他研究的結果。
英文摘要 Purpose: Alzheimer’s disease (AD) is a progressive neurodegenerative disease, and the structure of hippocampus could suffer from great loss of volume within the brains of AD patients. This study proposes a method by using T1-weighted images of hippocampus from ADNI database with 3D convolutional neural network to classify the patient of AD from normal subjects. The dataset included 420 images, all came from different subjects, in which the numbers of AD and CN were equal. As CNN classified subjects by the features of the input images, this study also tried to understand what kinds of features could influence the classification of AD.

Method: The preprocessing process was performed by the FreeSurfer software. The preprocessing included rescaling, intensity normalization, computing affine transformation to MNI305 atlas, and segmentation of structures. Then the hippocampal images were extracted. Data augmentation was performed by rotation. A classification of support vector machine was used as an accuracy reference. The purposed CNN contained 4 convolutional blocks, and networks of the left side and right side were concatenated after convolution layers. Six different models of CNN with two types of image input were trained. In addition to this basic model, the data of hippocampal sizes and ages of subjects were added as extra features into other CNN architectures with the different input data types. The first type of input images only contains the hippocampus, and the second type contains two channels of images: the hippocampus with its surrounding volume and the hippocampal mask. Cross validation was performed by k-fold validation with k=5. After training of CNNs, the backtracking of features took place. First, important factors at the latter layer of the CNN were found. The contribution of voxels in feature map of each convolution layers were then backtracked, generating the contribution maps of the input image. Finally, those contribution maps were compared with the subfields of hippocampus, to understand the influence of each subfields in the classification.

Result: The validation accuracies of different models and types of input were ranging from 0.79 to 0.85. The model having SVM-trained additional data that added with CNN output and using images with only hippocampus as input, showed the highest validation accuracy of 0.85 with a standard deviation of 0.03. The testing accuracies of different models and types of input were ranging from 0.78 to 0.85. The model having additional data concatenated with CNN output and using two channel input images, showed the highest testing accuracy of 0.85 with a standard deviation of 0.03. The backtracked result of the CNN showed that the subfield of CA1 and subiculum have significant influence over classification of AD.

Conclusion: The comparison between different models shows that additional input of ages and hippocampal size had no significant effect in improving classification results. For models with higher initial accuracy, overfitting is a clear problem. The backtracked resulted that indicated CA1 and subiculum have significant influence in the classification of AD could be exemplified by other researches.
論文目次 Abstract i
摘要 iii
誌謝 iv
Acknowledgement v
Index vi
1. Introduction 1
2. Method and Material 3
2-1. Data Set 3
2-2. Steps of the Method 3
2-3. Image Preprocessing 4
2-3-1. FreeSurfer image processing 4
2-3-2. Extracting Hippocampus image 4
2-3-3. Data Augmentation 6
2-3-4. Subfields of Hippocampus 7
2-4. Support Vector Machine Classification 7
2-5. 3D CNN 8
2-5-1. Image Convolution 8
2-5-2. Concatenation of Left and Right side 10
2-5-3. Label 11
2-5-4. Loss Function and Optimizer 11
2-5-5. Additional Information for CNN 11
2-5-6. Combined CNN and SVM 12
2-5-7. Weights Initialized by Trained CNN 14
2-5-8. Cross Validation 16
2-5-9. Tuning of Initialization of Learning Rates and Weights 16
2-6. Backtracking Critical Features of the 3D CNN 16
2-6-1. Computation of Import Factors in Concatenated Layer 16
2-6-2. Backtracked to Convolution Layer from the Important Factors 19
2-6-3. Backtracking between the Convolution Layers 20
2-6-4. Overlay of Contribution Maps 24
2-6-5. Comparison with Hippocampus Subfields 24
2-7. Experiment Environment 25
3. Result 26
3-1. Support Vector Machine 26
3-2. Convolutional Neural Network 26
3-2-1. CNN with only image input 26
3-2-2. Additional Information for CNN 28
3-2-3. Combined CNN and SVM 29
3-2-4. Weights Initialized by Trained CNN 31
3-2-5. Result of Testing Data 33
3-2-6. Tuning of Initialization of Learning Rates and Weights 34
3-3 Backtracking of Convolution Layers 37
3-3-1. Contribution Maps of Image Input Layers 37
3-3-2. Comparison with Hippocampus Subfields 37
4. Discussion 40
4-1. Result of CNNs 40
4-1-1. Comparison of the Types of Input Image 40
4-1-2. Comparison of Models 40
4-1-3. Comparison of Different Training Parameters 42
4-2. Backtracking of Feature Maps 42
5. Conclusion 44
Reference 45
參考文獻 [1] Goedert M, Spillantini MG. A century of Alzheimer’s disease. Science 314:777-781. 2006.
[2] Szabo K, Hennerici M.G, editors. The Hippocampus in Clinical Neuroscience. Karger, Basel, 2014.
[3] Karas GB, Scheltens P, Rombouts SA, et al. Global and local gray matter loss in mild cognitive impairment and Alzheimer’s disease. Neuroimage 23:708-716. 2004.
[4] Chan, D. et al. Change in rates of cerebral atrophy over time in early-onset Alzheimer's disease: longitudinal MRI study. Lancet 362: 1121-1122. 2003.
[5] J. Zhang, C. Yu, G. Jiang, W. Liu, L. Tong. 3D texture analysis on MRI images of Alzheimer’ disease. Brain Imaging and Behavior. 6:61-69. 2020.
[6] Cuingnet R, Gerardin E, Tessieras J, Auzias G, Lehericy S, Habert MO, et al. Automatic classification of patients with Alzheimer’s disease from structural MRI: a comparison of ten methods using the ADNI database. Neuroimage. 56:766-81. 2010.
[7] M.A. Ebrahimighahnavieh, S. Luo, R. Chiong. Deep learning to detect alzheimer’s
disease from neuroimaging: A systematic literature review. Comput. Methods
Programs Biomed. 187. 2020.
[8] Liu M, Cheng D, Yan W. Classification of Alzheimer's disease by combination of convolutional and recurrent neural networks using FDG‐PET images. Front Neuroinform. 12(35): 1-36. 2018.
[9] D. Cheng, M. Liu. Combining convolutional and recurrent neural networks for Alzheimer's disease diagnosis using PET images. Proc. 2017 IEEE Int. Conf. Imag. Syst. Techn. 1-5. 2017.
[10]Wang SH, Phillips P, Sui Y, Liu B, Yang M, and Cheng H. Classification of alzheimer’s disease based on eight-layer convolutional neural network with leaky rectified linear unit and max pooling. J. Med. Syst. 42(5):85. 2018.
[11] S. Korolev, A. Safiullin, M. Belyaev and Y. Dodonova. Residual and plain convolutional neural networks for 3D brain MRI classification. Proc. IEEE 14th Int. Symp. Biomed. Imag. 10751: 835-838. 2017.
[12] Huang Y, Xu J, Zhou Y, Tong T, Zhuang X, Alzheimer’s Disease Neuroimaging Initiative (ADNI). Diagnosis of Alzheimer's disease via multi-modality 3D convolutional neural network. Front Neurosci. 13:509. 2019.
[13] H. Karasawa, C. Liu and H. Ohwada. Deep 3D Convolutional Neural Network Architectures for Alzheimers Disease Diagnosis. Intelligent Information and Database Systems. 287-296. 2018.
[14] Ruoxuan Cui, Manhua Liu. Hippocampus analysis based on 3D CNN for Alzheimer’s disease diagnosis. Tenth International Conference on Digital Image Processing (ICDIP 2018). vol.10806, p. 108065O. 2018.
[15] Mugler JP 3rd, Brookeman JR. Rapid three-dimensional T1-weighted MR imaging with the MP-RAGE sequence. J Magn Reson Imaging 1:561-7. 1991.
[16] R. G. Boyes, J. L. Gunter, C. Frost, A. L. Janke, T. Yeatman, D. L. G. Hill, et al. Intensity non-uniformity correction using N3 on 3-T scanners with multichannel phased array coils, Neuroimage. vol. 39, no. 4, pp. 1752-1762. 2008.
[17] Dale A, Fischl B, Sereno M. Cortical surface‐based analysis I: segmentation and surface reconstruction. Neuroimage 9: 179-194. 1999.
[18] Fischl B, Salat DH, Busa E, Albert M, Dieterich M, Haselgrove C, van der Kouwe A, Killiany R, Kennedy D, Klaveness S, Montillo A, Makris N, Rosen B, Dale AM. Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron. 33:341-355. 2002.
[19] Iglesias, J. E. et al. A computational atlas of the hippocampal formation using ex vivo, ultra-high resolution MRI: application to adaptive segmentation of in vivo MRI.Neuroimage. 115:117-137. 2015.
[20]Chih-Chung Chang, Chih-Jen Lin. LIBSVM : a library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2:27:1-27:27, 2011. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm
[21]Muller, R., Kornblith, S., and Hinton, G. E. When does label smoothing help? In Advances in Neural InformationProcessing Systems, pp. 4696-4705, 2019.
[22] Zeiler, M. D. ADADELTA: An adaptive learning rate method. arXiv:1212.5701. 2012.
[23]Parker TD, Cash DM, Lane CAS et al. Hippocampal subfield volumes and pre-clinical Alzheimer’s disease in 408 cognitively normal adults born in 1946. PLoS ONE. 14, e0224030. 2019.
[24] D.H. Adler et al. Characterizing the human hippocampus in aging and alzheimers disease using a computational atlas derived from ex vivo MRI and histology
Proc. Natl. Acad. Sci. Unit. States Am. 115: 4252-4257. 2018.
[25] Carlesimo, GA, Piras, F, Orfei, MD. Atrophy of presubiculum and subiculum is the earliest hippocampal anatomical marker of Alzheimer’s disease. Alzheimer Dement Diagnos 1: 24-32. 2015.
論文全文使用權限
  • 同意授權校內瀏覽/列印電子全文服務,於2024-01-11起公開。
  • 同意授權校外瀏覽/列印電子全文服務,於2024-01-11起公開。


  • 如您有疑問,請聯絡圖書館
    聯絡電話:(06)2757575#65773
    聯絡E-mail:etds@email.ncku.edu.tw