透過您的圖書館登入
IP:3.83.87.94
  • 學位論文

探討機器與深度學習於99mTc-TRODAT-1 SPECT影像分類巴金森氏症多期別之表現

Investigation of Classified Multiple Stages of Parkinson Disease by Machine Learning and Deep Learning on 99mTc-TRODAT-1 SPECT Images

指導教授 : 陳泰賓
共同指導教授 : 杜維昌(Wei-Chang Du)
本文將於2025/07/03開放下載。若您希望在開放下載時收到通知,可將文章加入收藏

摘要


單光子電腦斷層影像(Single Photon Emission Computed Tomography, SPECT)用於臨床帕金森氏症(Parkinson Disease, PD)患者具有高臨床檢出率。然而,PD患者SPECT影像分析大多採用感興趣區域選取(Region of Interest, ROI)方式,因此受限ROI大小,而導致PD多期別分類效能受限。故本研究設計全腦(去除背景區域)活性分布與三維紋狀體活性體積特徵值萃取方法,經由機器及深度學習方法建立多期別PD分類模型。 採用回溯性分組實驗設計,收集99mTc-TRODAT-1顯影劑(對比劑藥物)進行腦部SPECT造影,成功收集202筆資料(年齡分佈為25至91歲平均69歲)。接著依據Hoehn and Yahr Scale (HYS) 標準將PD患者分為正常(n=6)、早期(HYS I~II, n=22, 27)中期(HYS III, n=53)與晚期(HYS IV~V, n=87, 7)。針對SPECT影像結合單一域值法及三維區域種子成長演算法,定義六種SPECT影像特徵值;分別為Skewness、Kurtosis、Cyhelsky's Skewness Coefficient、Pearson's Median Skewness、Dopamine Transporter Activity Volume (DTAV)、Dopamine Transporter Activity Maximum;接著使用Logistic Regression (LR)與Support Vector Machine (SVM)做為PD期別分類模型,採用2-Fold交叉驗證方式進行模型評估。深度學習演算法採用六種卷積神經網路(Convolutional Neural Network, CNN),包括AlexNet、GoogLeNet、Residual Neural Network、VGG、DenseNet 與三維CNN模型,其中影像輸入層(imageInputLayer)包括有2D灰階、2D彩色及3D影像三種方式。每位受試者取得紋狀體最大活性切片處(User Define)與前後各兩張,共計5張影像建立影像資料集,2D灰階、2D彩色及3D影像總影像張數別為1010、1010及202;採取資料集之70%與30%進行訓練及驗證。效能評估方式均採用準確度、靈敏度、陽性預測率、F-score以及Kappa一致性等指標。 研究結果顯示透過顯著特徵值Skewness、Kurtosis及DTAV,利用LR建立PD四期分類模型較SVM佳,其準確度、靈敏度、陽性預測率、F-score以及Kappa一致性分別為0.71、0.88、0.78、0.83、0.54。六種卷積神經網路建立四分類模型,在以2D灰階、2D彩色及3D影像為分類基礎之模型最佳者,分別為AlexNet、DenseNet201及3D CNN具最好分類準確度0.83、0.85及0.66。針對六分類模型在以2D灰階、2D彩色及3D影像為分類基礎之模型最佳者,分別為VGG19、DenseNet201及3D CNN具最好分類準確度0.78、0.78及0.53。 本究結果顯示CNN建立之分類模型準確度比機器學習方法高,其對PD SPECT影像進行四及六分類準確率達85%及78%。然而,機器學習方法透過有限影像特徵建立模型時間花費較少且影像特徵具有可解釋性及臨床意義;反之,CNN建立分類模型具有費時高、參數設定多、運算過程產生的中間結果其可解釋性低或不具有臨床意義。

並列摘要


Single Photon Emission Computed Tomography (SPECT) have been employed to detect the stages of Parkinson disease (PD). However, the region of interest (ROI) base was mostly used to analysis of SPECT PD image. Due to limited by the size of ROI especially in the multi-stage classification of PD. Therefore, this study designed a feature extraction method with whole brain (remove background area) active distribution and three dimensional of striatum active volume. At the same time used machine learning and deep learning methods to establish a multiple stages classification model of PD. In the retrospective study, the 99mTc- TRODAT-1 was used for brain SPECT imaging. A total of 202 cases were collected (age distribution from 25 to 91 years old with an average of 69 years old). Then, according to the Hoehn and Yahr Scale (HYS) standard, PD patients were divided into early (HYS I~II), mid (HYS III) and late (HYS IV~V) stage. In PD analysis, a single threshold and seed region growing algorithm were used to calculate a total of six features. There were Skewness, Kurtosis, Cyhelsky's Skewness Coefficient, Pearson's Median Skewness, Dopamine Transporter Activity Volume (DTAV), and Dopamine Transporter Activity Maximum. And using logistic regression (LR) and support vector machine (SVM) methods to establish the classification models for PD stage. Then, using 2-fold cross validation method to evaluate the model. In deep learning, it compares with six convolution neural network (CNN). There were AlexNet, GoogLeNet, Residual Neural Network, VGG, DenseNet, and 3D CNN model. The image input layer included 2D grayscale, 2D color, and 3D image types. Each subject obtains the maximum active slice of the striatum (User Define) and gets the previous and the next two images of the active slice. The total number of images in 2D gray, 2D color and 3D images were 1010, 1010 and 202, respectively. The training and validate sets were 70% and 30%. The accuracy, recall, precision, F-score and Kappa values were used to evaluate the performance of models. In result, the study use of LR to establish the classification model through Skewness, Kurtosis and DTAV significance features was better than SVM. The accuracy, recall, precision, F-score and Kappa values were 0.71, 0.88, 0.78, 0.83, and 0.54 with respectively. Six types of CNN were used to classify four categories. The best models based on 2D grayscale, 2D color and 3D images were AlexNet, DenseNet201 and 3D CNN. The accuracy were 0.83, 0.85 and 0.66, respectively. For six categories, the best models based on 2D grayscale, 2D color and 3D images were VGG19, DenseNet201 and 3D CNN. The accuracy were 0.78, 0.78 and 0.53, respectively. The results of this study shown that the accuracy of the classification model established by CNN was higher than machine learning methods. In four and six multiple stages of PD SPECT images were 85% and 78%. However, the machine learning method cost less time to build a model and the image features had interpretable and clinically significant. On the contrary, the CNN establishment classification model had high time-consuming, parameter setting, and the results had low interpretability and no any clinical meaningful.

參考文獻


[1] T. Popa, L. Ibanez, E. Levy, A. White, J. Bruno, and K. Cleary, "Tumor volume measurement and volume measurement comparison plug-ins for VolView using ITK," in Medical Imaging, pp. 61411B-61411B-8, 2006.
[2] A. Dutour, J. Monteil, F. Paraf, J. L. Charissoux, C. Kaletta, B. Sauer, et al., "Endostatin cDNA/cationic liposome complexes as a promising therapy to prevent lung metastases in osteosarcoma: study in a human-like rat orthotopic tumor," Molecular Therapy, vol. 11, pp. 311-319, 2005.
[3] H. Sun, "An improved positron emission tomography (PET) reconstruction of 2D activity distribution using higher order scattered data," 2016.
[4] D. L. Bailey and K. P. Willowson, "An evidence-based review of quantitative SPECT imaging and potential clinical applications," J Nucl Med, vol. 54, pp. 83-9, 2013.
[5] A. Galvan, A. Devergnas, and T. Wichmann, "Alterations in neuronal activity in basal ganglia-thalamocortical circuits in the parkinsonian state," Frontiers in neuroanatomy, vol. 9, p. 5, 2015.

延伸閱讀