應用類神經網路於帕金森氏症SPECT腦血流灌注影像之辨識

Using Artificial Neural Network to identify SPECT Brain Perfusion Imaging of Parkinson's Disease

帕金森氏症、中風與癡呆併列為老年人三大疾病，影響國人健康甚大。而Tc-99m HMPAO放射性藥劑配合上單光子電腦斷層掃描（single photon emission computed tomography, SPECT）的腦部灌注影像已被經常地應用在診斷帕金森氏症。傳統醫師在做腦部SPECT影像診斷時，需要人工圈選感興趣區（ROI, Region Of Interest），然後使用半定量分析測量其計數值，之後使用統計軟體處理實驗數據。但是因不同人圈選之位置、大小不同，會造成人為統計誤差，使診斷結果不一。本研究的目的在於設計出以類神經網路（ANN, Artificial Neural Network）為基礎之電腦輔助診斷模型，其中以統計參數對映圖示（SPM, Statistical Parametric Mapping）輔助決定圈選ROI與將影像半定量分析的數值輸入類神經網路進行訓練，然後建立出較為客觀的用於分析SPECT腦灌注影像的模型作為一種輔助診斷方法，以協助醫師做帕金森氏症的鑑別診斷。而訓練出的類神經網路模型ROC曲線下面積為0.978，整體敏感度為95%、特異度為100%與準確度達97%，可見此電腦輔助診斷模型可有效的區分帕金森氏症與正常人的SPECT腦部灌注影像。

Parkinson's disease, stroke and dementia are the three major diseases of elderly population. SPECT (single photon emission computed tomography) with the Tc-99m HMPAO (technetium-99m hexamethylpropyleneamine oxime) is frequently used in diagnosis of Parkinson's disease. The classical approach is to select region of interest (ROI) manually. However, manual approaches are biased and take time and low efficiency. In this study an Artificial Neural Network (ANN) based computer-aided diagnosis model is proposed for diagnosis of Parkinson's disease. In results, the trained ANN model had an overall performance that the area under curve (AUC) of the receiver operating characteristic (ROC) curve is 0.978, sensitivity is 95%, specificity is 100% and accuracy is 97%. The result revealed that this computer aid diagnosis model can effectively identify SPECT brain perfusion images between Parkinson's disease and normal.