Diagnosis of Alzheimer disease from PET images by deep learning technique

The medical condition that develops in the form of memory loss, dementia (dementia) and decreased cognitive functions due to the death of brain cells over time is called Alzheimer's disease (AD). Although no drug has been found for its treatment yet, there are successful results that may lead to the understanding of its cause, development, and solutions. Examination of the information about the presence and progression of AD with the help of deep learning architectures through brain images can increase diagnostic accuracy. Some studies have been conducted to develop and validate deep learning algorithms to predict the final diagnosis of AD, mild cognitive impairment, or brain Fluoro-deoxy-glucose (FDG)-PET, and to compare its performance with that of radiological readers. In this study, with the help of deep learning, diagnosis of AD was attempted by adapting transfer learning technique and DenseNet121, InceptionV3, MobileNet, MobileNetV2, ResNet50, ResNet101, VGG16 and Xception architectures over FDG-PET images, and the condition of the brain tissue affected by the disease was investigated. When the findings of the study were examined, Xception (99%) and DenseNet121 (99%) architectures yielded the highest training accuracy, precision, sensitivity and F score. Considering the test accuracy, precision, sensitivity and F1 score obtained as a result of training, Xception (100%) and DenseNet121 (100%) architectures came to the fore. Among the models used, Xception, VGG16 and ResNet101 took the longest training and test durations, while MobileNet and MobileNetV2 models took the shortest time. As a result of training with different data sets, it was seen that the success of the model increased as the amount of data increased. The study showed the performance of transfer learning architectures in diagnosing AD over brain PET images and suggested that it can be studied using larger datasets on MR images with architectures that show high success according to the training results.