• 系統編號 RN9704-0424
• 計畫中文名稱 發展一個以獨立元件分析為基礎之自動分類器用以評估巴金森氏病的臨床診斷
• 計畫英文名稱 Development and Evaluation of an Ica-Based Automatic Classifier for Parkinson's Disease Diagnosis Using Spect Data
• 主管機關 行政院國家科學委員會 • 計畫編號 NSC95-2314-B038-067
• 執行機構 台北醫學大學醫學資訊研究所
• 本期期間 9508 ~ 9607
• 報告頁數 16 頁 • 使用語言 中文
• 研究人員 徐建業; 邱泓文; 徐榮隆 Hsu, Chien-Yeh
• 中文關鍵字 醫學影像; 獨立元件分析; 巴金森氏病; 認知功能; 單光子電腦斷層攝影; 診斷輔助; 自動分類
• 英文關鍵字 SPECT; Parkinson's disease; Regional cerebral blood flow; Independent component analysis; Statistical parametric mapping
• 中文摘要 中文摘要:在過去二十年間，生物醫學及工程上最重要的進展之一，就是在非侵入式記錄(non-invasive recording)技術的突飛猛進。這樣的技術讓我
們能設計、從事各種複雜的實驗，記錄高精密度的腦電波、心電圖、 肌電圖、 磁振照影及功能性磁振照影。從龐大複雜的數據中，擷取出有意義的
資訊。目前相當受矚目的『獨立元件分析』(Independent ComponentAnalysis, ICA)演算法，分析功能性腦照影訊號，可將多管道(concurrentmultiple- channel)生醫訊號分解成完全獨立的訊號(independent component)，並進一步探討這些獨立訊號與實驗行為(task behavior and performance) 的相關 性。這個分析方法在過去幾年已經應用到各種基礎腦神經科學的數據分析。巴金森氏病(PD)是老年常見的神經退化性疾病，其主要的症狀為動作遲 緩、僵直、姿態不穩與顫抖。雖然經過數十年的研究，對於巴金森氏病的腦血流變化仍然未有定論，而目前的單光子電腦斷層攝影對於評估局部的 腦血流與了解疾病的病理生理學是一項有用的工具。我們在去年的計畫中已經成功的運用此分析方式（獨立元件分析）來比較正常人與巴金森氏 病人其大腦中的局腦血流差異。並且對於巴金森氏症大腦局腦血流異常的病人，比較其在不同大腦區域之間的關係。在過去兩年的計畫中(NSC93- 2320-B-038 –033) 和(NSC-94-2320-B-038-015)，我們已經利用獨立元件分析的方法開發出一套演算法來分析正常人與巴金森氏病人的單光子電腦
斷層攝影影像資料，並比較正常人與巴金森氏病人其大腦中的局部腦血流差異的位置。我們的結果顯示，辨識出的PD 相關位置與 DeLong's PD
model 的預測相當吻合。同時我們更進一步運用機械學習理論的方法，將由獨立元件分析所得到的巴金森氏症大腦局腦血流異常區域當成特殊指 標，運用Support VectorMachine(SVM)自動分類器(AutomaticClassifier)依此指標來做資料的分類。我們的研究結果顯示，運用此一完整的方式可以 達到92%的敏感度與 96%的特異度，其 Receiver OperatorCharacteristics(ROCs) curves 下面積更達 99%，顯見這方法可以有效的協助臨床醫師做診 斷上的協助指標。本計劃之主要目的，在延續目前之研究結果，並建立一個以統計分析為基礎之自動分類器用以輔助及評估巴金森氏病的診斷平
台，以應用於臨床上。我們將收集更多的資料以進行此方式的穩定性並使得統計更有一致及完整性。同時，病患ICA 元件之權重與其臨床檢驗檢
查分數之相關性也將被進一步的探討，可以幫助我們了解rCBF 的改變是否與 PD 之病程有關係。另一方面，本計劃亦將發展並評估用於診斷巴金 森氏症的自動分類器(Automatic Classifier)其參數的選擇與訂定，將有助於未來對其他類似巴金森氏症的疾病區別。本研究之結果將對 PD 之臨床 研究與自動分析有重要的影響。此外，藉由本計劃的研究方法與成果，將有助於應用到其他疾病的診斷分析上。
• 英文摘要
This study investigates regional cerebral blood flow (rCBF) changes in patients with Parkinson's disease using independent component analysis (ICA) followed by statistical parametric mapping (SPM). Methods: 99mTchexaemethyl-propyleneamine oxime (99mTc-HMPAO) was used as the CBF tracer for rCBF measurements. A single photon emission computerized tomography (SPECT) study was performed on 62 patients with Parkinson's disease in various disease stages, and also on 51 aged-matched controls. SPECT images were first spatially normalized to standard space, concatenated, and then subjected to ICA decomposition. The resulting image components were then separated by logistic regression into two sets: disease-related components, whose subject weights differenced between groups and non-disease related components, whose subject weights exhibited no group difference. Components of each set were back-projected and summed across components. The resultant rCBF images were normalized to the global CBF for each subject and then analyzed using SPM to compare the rCBF values changes between Parkinson's disease and control subject. Results: In the disease-related image subspace, patients with Parkinson's disease exhibited significantly higher adjusted rCBF in the subthalamic nucleus, putamen, globus pallidum, thalamus, brainstem, and anterior lobe of cerebellum, and significant hypoperfusion in the supplementary motor plus dorsolateral prefrontal, pariteo-occipital cortex, insula, and cingulate gyrus. In the non-disease related image subspace, very few regions showed a significant group difference. Using SPM only without ICA separation gave significantly lower peak t value and at a smaller number of image voxels. Some of the regions revealed by ICA to be affected by Parkinson's disease have not shown significant changes in previous HMPAO-SPECT studies, though those are central to the pathophysiological model of Parkinson's disease. Finally, SVM could correctly classified normal controls from disease patient. Conclusion: In a HMPAO-SPECT study, ICA-based separation of normalized images into disease-related and unrelated subspaces revealed more disease-related brain regions than applying SPM directly. The diseased-related regions indicated by ICA are consistent with the current model of pathophysiology in Parkinson's disease, though their rCBF changes in Parkinson's disease have not been fully demonstrated by any previous single functional imaging study. Also, the SVM classification method could help clinical doctor diagnosis between PD and controls. Thus ICA combine with SVM method provides a new and more comprehensive method for testing functional and brain circuit models in Parkinson's disease.
