透過您的圖書館登入
IP:3.91.43.22
  • 期刊

以少量ECD單光子斷層影像實現深度學習模型預測阿茲海默症

Deep Learning Model to Predict Alzheimer's Disease Using a Small Number of Tc-99m-ECD SPECT Images

摘要


目的:本研究將科技部補助核能研究所所收集建置之「國人專屬之核醫腦功能影像資料庫」中的部分鎝-99m ethyl cysteinate dimer(ECD)單光子斷層影像(single photon emission tomography, SPECT)造影數據進行深度學習模型訓練,研究以少量ECD SPECT影像預測阿茲海默症(Alzheimer's disease, AD)的可能性。方法:首先將101位AD和86位認知功能正常者(normal cognition, NC)的鎝-99m-ECD SPECT影像,各自降維成3張維度為380 × 380的二維影像,採用影像視覺領域中的ResNet50模型及其預訓練結果,透過氟-18氟代去氧葡萄糖正子造影影像預訓練以學習核醫影像之特徵,最後藉由遷移學習將前述學習之訓練權重應用於本研究所使用之鎝-99m-ECD SPECT影像訓練上。以訓練曲線及特徵視覺化之結果評估訓練情況,進一步以各種性能指標評估AD與NC分辨之成效。結果:在常規硬體配備環境下,透過兩階段的遷移學習技術,成功訓練鎝-99m-ECD SPECT影像分辨AD及NC。直接以ResNet50模型訓練之靈敏度與特異性分別為78.3%及53%;ResNet50模型結合預訓練參數之靈敏度與特異性分別為82.6%及66.6%,證實使用預訓練及遷移學習之效能提升明顯。結論:本研究提供一個務實的方式,能快速針對手邊擁有的少量核醫影像數據,以人工智慧進行自動影像特徵萃取,協助分析及探究神經退化性疾病的認知功能障礙與生物標記之關聯。

並列摘要


Background: In this study, based on the "Taiwanese Nuclear Medicine Brain Functional Image Database" built by Institute of Nuclear Energy Research and subsidized from the Ministry of Science and Technology, part of the Tc- 99m- ethyl cysteinate dimer (ECD) single photon emission tomography (SPECT) images were used for deep learning model training. The goal is to investigate the feasibility of using a small number of ECD SPECT images to predict the diagnosis of Alzheimer’s disease (AD). Methods: First, every Tc-99m-ECD SPECT image (101AD and 86 normal cognition [NC]) was dimension reduced to three sets of two-dimensional images with dimensions of 380 × 380. The ResNet50 model in the field of image vision and its pre-trained results were used, further through fluorodeoxyglucose (FDG) PET images pre-training to learn the characteristics of nuclear medicine images, finally transfer learning to apply on Tc-99m-ECD SPECT images. Evaluate the training situation with the results of the training curve and feature visualization, and further evaluate the ability to identify AD and NC by various performance indexes. Results: Based on a two-stage transfer learning technique, and using Tc-99m-ECD SPECT images, a deep learning model was successfully trained to distinguish AD from NC using a piece of conventional computing equipment. The FDG images used in the pre-trained model were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The sensitivity and specificity of training directly with the ResNet50 model was 78.3% and 53%, respectively; the sensitivity and specificity of the ResNet50 model combining the pre-trained parameters was 82.6% and 66.6%, respectively. The result demonstrates that the performance of an artificial intelligence (AI) model on ECD SPECT images for AD diagnosis using pre-training and transfer learning is significantly improved. Conclusions: This study provides a pragmatic way to quickly extract the diagnostic information from a small amount of nuclear medicine imaging data at hand, and applies an AI model to perform automatic image feature extraction to assist in analyzing and exploring the correlation between cognitive dysfunction and biomarkers in neurodegenerative diseases.

延伸閱讀