卷積神經網路技術輔助診斷失智症腦部灌注影像

Evaluation of Cerebral Perfusion in Tc-99m ECD SPECT Imaging by Convolutional Neural Network: Dementia Diseases Clinical Diagnosis Auxiliary

陳瑞仁(Jui-Jen Chen) ； 張簡鴻年(Hung-Nien Chang Chien) ； 顏宏旗(Hung-Chi Yen) ； 張雁翔(Yen-Hsiang Chang)

《北市醫學雜誌》 預刊文章 Pp. 1-28-0090

https://doi.org/10.6200/TCMJ.202201/PP.0023

摘要

目的：本研究主要收集高雄長庚紀念醫院Tc-99m ECD SPECT（Ethyl Cysteinate Dimer, Single Photon Emission Computed Tomogrphy）腦部灌注影像，利用卷積神經網路技術對失智症腦部灌注影像進行分析，SPECT臨床影像中，由腦部血流分佈擷取出最重要的特徵。方法：依結構化三階段萃取規則，共收集153例病歷，資料預處理以醫師平時經驗常判讀的區堿進行提取，去除周圍影像雜訊經降噪後，通過CNN（Convolutional Neural Network）模型對病人是否患有失智症影像進行訓練，並萃取影像中顯著的特徵值進行分類分析。結果：預測準確率、敏感性和特異性分別為89.7%、92.9%和90.9%，ROC（Area Under Curve）各分類器對本研究模型評估，AUC下面積大於0.9小於1.0，有極佳的鑑別力，故本研究模型辨識準確率高且具有預測價值之模型。結論：此輔助系統將大大減少醫生影像判讀和分析所需的時間，並協助臨床專家更準確診斷腦部影像問題。

關鍵字

失智症 ； 腦部灌注影像 ； 機械學習 ； 卷積神經網路 ； 核子醫學

並列摘要

Objective: This study primarily analyzed the Tc-99m ECD SPECT (Ethyl Cysteinate Dimer, Single Photon Emission Computed Tomography) Cerebral perfusion images collected from Kaohsiung Chang Gung Memorial Hospital using Convolutional Neural Network (CNN) technology. The most important features of blood flow distributions in the brain were extracted. Methods: According to the structured three-stage extraction rule, 153 cases were collected, and the data were preprocessed and extracted from the areas often interpreted by physicians. After removing the surrounding noises and noise reduction, dementia and non-dementia images were used in the training of the CNN model, and significant features in the images were extracted, categorized, and analyzed. Results: The precision, sensitivity, and specificity of the model were 89.7%, 92.9%, and 90.9%. The area under the curve was larger than 0.9 and smaller than 1.0, indicating an excellent differentiation ability. Thus, the model has high precision and predictive value in target identification. Conclusions: The auxiliary system will greatly reduce the time required for physicians to judge and analyze the images, and help clinical experts to diagnose the brain image even more preciously.

並列關鍵字

dementia diseases ； cerebral perfusion image ； machine learning ； convolutional neural network ； nuclear medicine