請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/72432
標題: | 以多模態深度學習網路進行帕金森氏症磁振影像之電腦輔助診斷 Deep Neural Networks for Computer Aided Diagnosis of Parkinson’s Disease Using Multimodal MR Images |
作者: | Yu-Hsueh Wu 吳侑學 |
指導教授: | 陳右穎(You-Yin Chen) |
關鍵字: | 電腦輔助診斷,帕金森氏症,統一帕金森病評定量表,深度學習,類神經網路, Computer Aided Diagnosis,Parkinson’s Disease,Deep Learning,Neural Network, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 帕金森氏症的電腦輔助診斷一直都是一項重要的研究議題,好的電腦輔助診斷系統能輸出正確的預測診斷結果,不僅可以加速醫師診斷的效率,還可以為臨床診斷減少人為的疏失,提升品質。
本論文基於傳統的電腦輔助診斷並加以精煉,提出了能同時輸出「預測診斷」以及產生統一帕金森病評定量表(Unified Parkinson's Disease Rating Scale Scores ,UPDRS)的系統。由於臨床量表對於帕金森氏症的診斷來說,是最為耗時卻又最為準確的醫療資訊,本研究能大大的節省醫師、病患以及醫療人員的時間,同時又能提供給醫師更豐富的醫療資訊做為診斷基礎。除此之外整個系統採用多模態輸入的模型,因此整個電腦輔助診斷的過程能夠綜合多種臨床資訊並做出更全面、更有可信度的預測診斷。 本論文透過類神經網路來開發此電腦輔助診斷系統,並且設計了主要由生成器(Generator)和分類器(Classifier)構成的兩段式結構模型。生成器綜合不同模態的原始臨床資訊,做出判斷並生成預測臨床量表;分類器則根據生成器輸出的臨床量表來分類受試者,輸出預測診斷結果。本論文以全腦的T1加權影像與靜息態功能性磁振影像,做為訓練模型的原始臨床資訊,在進入類神經網路的訓練階段之前,我們會先針對兩者做以腦區為基礎的資料前處理,T1加權影像透過Freesurfer分析,抽取出各個腦區的特徵值,髓質的部分包括體素數目、體積、強度平均值、強度標準差、強度最小值、強度最大值和強度範圍七項;皮質的部分包括頂點數目、表面積、體積、厚度平均值、厚度標準差、平均曲率和高斯曲率這七項;靜息態功能性磁振影像則透過FSL和AFNI兩種函式庫來分析,每個腦區會計算出三種特徵值,分別是fALFF、ICA maps和ReHo。前處理過後的特徵值會經過一層投影層,目的是將不同向量空間的輸入模態投影到相同的向量空間。接著,投影過後的特徵值會輸入生成器中,生成器又可分為兩個部分,分別是Attention Layer 和Highway Network,前者目的是讓模型去主動學習對於診斷相對重要的腦區,後者則負責生成預測臨床量表,它是一種類神經網路的變形,目的是讓生成器的網路結構可以堆疊的更深。最後,由簡單類神經網路構成的分類器,會根據生成的預測臨床量表做出預測診斷,將受試者分類成「僵直病患」、「震顫病患」以及「健康受試者」三類。 網路的訓練數據由45位受試者構成,包含了26名帕金森氏症患者和19名健康受試者,26名患者又可依症狀分為僵直組20人和震顫組6人。 模型的訓練與測試採用了三次交叉驗證的方法,在預測診斷的正確率上達到了91%的正確率,並在預測臨床量表與真實臨床量表的相關係數上達到0.92的高度相關。此外,本論文對於模型進行了剝離測試,驗證了模型中的各個組件和多模態輸入都有其必要性,能夠幫助模型有更好的表現;同時,我們透過層數測試的實驗,證實了類神經網路的深度直接影響了預測診斷的正確率,層數越深,正確率越高。 本論文針對帕金森氏症,建立了在預測診斷結果有高準確度、在預測臨床量表有高相似性之電腦輔助診斷系統,此系統能基於多模態的輸入資訊做出更完善的輔助診斷,並且模態的數目與種類具有極高的可擴充性。 Computer aided diagnosis(CAD) is very important for Parkinson’s Disease since CAD can not only generate accurate diagnostic prediction but improve quality of diagnosis. In this study, we proposed a refined CAD system. Our system generated both diagnostic prediction and predicted Unified Parkinson's Disease Rating Scale (UPDRS) scores in order to provide doctors with more medical information and made diagnosis more efficient. Furthermore, the whole system was designed as a multimodal model which made the decision process more comprehensive. We implemented a refined CAD system by proposing an innovative two-staged neural network model composed by the Generator and the Classifier. The Generator integrated medical information from different modalities and generated predicted clinical scales. The Classifier made diagnostic prediction based on predicted clinical scales from the Generator. We used whole brain T1 weighted images and resting state functional MRI as the raw data to train our model. But before training phase, we carried out a ROI based preprocessing. T1 weighted images were preprocessed using Freesurfer. Freesurfer extracted features from each ROI. For subcortex, the features included voxel number, volume, average intensity, standard deviation of intensity, minimum intensity, maximum intensity, intensity range; for cortex, the features included vertex number, surface area, volume, average thickness, standard of thickness, average curvature, Gaussian curvature. Resting state functional MRI were preprocessed using FSL and AFNI library. Each ROI contained 3 different features including fALFF, ICA maps and ReHo. After preprocessing, the Projection Layer would project features from different modalities into the same vector space. Then, the Generator would generate predicted clinical scales based on those projected features. The Generator was composed of two part, the Attention Layer and the Highway Network. The Attention Layer focused on learning the importance of each ROI for diagnosis; the Highway Network was a transformation of neural network specialized in stacking network deeper, and its main target was to generate predicted clinical scales. At last, the Classifier, which was actually a basic neural network structure, would make diagnostic prediction based on predicted clinical scales, and classified subjects into 3 types: PD Tremor, PD Rigidity and health control. The dataset we used in this paper contained 45 subjects including 26 Parkinson’s disease patients and 19 health control subjects. Also, we separated 26 patients into two groups, PD Tremor and PD Rigidity, consisting of 6 and 20 patients respectively. We trained and tested our model using 3-fold cross validation. Our model reached 91% accuracy on diagnostic prediction, and the correlation coefficient between real clinical scales and predicted clinical scales achieved 0.92. Moreover, the ablation test we carried out on this model proved that each component we designed, as well as the multimodal input mechanism, could improve our model. And the layer test proved that the deeper the neural network was, the better our model performed. As a result, we successfully build up a multimodal CAD system for Parkinson’s disease with high accuracy on diagnostic prediction and high correlation on predicted clinical scales. Moreover, the scalability of input modality was unlimited. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/72432 |
DOI: | 10.6342/NTU201803387 |
全文授權: | 有償授權 |
顯示於系所單位: | 醫學工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf 目前未授權公開取用 | 2.4 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。