The correct differential diagnosis of dementia has an important impact on patient treatment and follow-up care strategies. Tc-99m-ECD SPECT imaging, which is low cost and accessible in general clinics, is used to identify the two common types of dementia, Alzheimer’s disease (AD) and Lewy body dementia (LBD). Two-stage transfer learning technology and reducing model complexity based on the ResNet-50 model were performed using the ImageNet data set and ADNI database. To improve training accuracy, the three-dimensional image was reorganized into three sets of two-dimensional images for data augmentation and ensemble learning, then the performance of various deep learning models for Tc-99m-ECD SPECT images to distinguish AD/normal cognition (NC), LBD/NC, and AD/LBD were investigated. In the AD/NC, LBD/NC, and AD/LBD tasks, the AUC values were around 0.94, 0.95, and 0.74, regardless of training models, with an accuracy of 90%, 87%, and 71%, and F1 scores of 89%, 86%, and 76% in the best cases. The use of transfer learning and a modified model resulted in better prediction results, increasing the accuracy by 32% for AD/NC. The proposed method is practical and could rapidly utilize a deep learning model to automatically extract image features based on a small number of SPECT brain perfusion images in general clinics to objectively distinguish AD and LBD.
Alzheimer’s disease (AD) and vascular dementia (VaD) are the two most common forms of dementia. However, their neuropsychological and pathological features often overlap, making it difficult to distinguish between AD and VaD. In addition to clinical consultation and laboratory examinations, clinical dementia diagnosis in Taiwan will also include Tc-99m-ECD SPECT imaging examination. Through machine learning and deep learning technology, we explored the feasibility of using the above clinical practice data to distinguish AD and VaD. We used the physiological data (33 features) and Tc-99m-ECD SPECT images of 112 AD patients and 85 VaD patients in the Taiwanese Nuclear Medicine Brain Image Database to train the classification model. The results, after filtering by the number of SVM RFE 5-fold features, show that the average accuracy of physiological data in distinguishing AD/VaD is 81.22% and the AUC is 0.836; the average accuracy of training images using the Inception V3 model is 85% and the AUC is 0.95. Finally, Grad-CAM heatmap was used to visualize the areas of concern of the model and compared with the SPM analysis method to further understand the differences. This research method can quickly use machine learning and deep learning models to automatically extract image features based on a small amount of general clinical data to objectively distinguish AD and VaD.
24 March 2025 in Journal of Medical and Biological Engineering