Effect of Data Augmentation of F-18-Florbetaben Positron-Emission Tomography Images by Using Deep Learning Convolutional Neural Network Architecture for Amyloid Positive Patients

Abstract

Early diagnosis of dementia helps in finding suitable treatments that reduce or even prevent future cognitive dysfunction of patients. In this paper, we use a convolutional neural network to classify the brain positron-emission tomography (PET) image of Alzheimer’s disease (AD) and mild cognitive impairment (MCI) in a normal control (NC). The purpose of this study is to investigate the influence of the data increment method and the number of data for the best accuracy in the convolutional neural network (CNN) by using the AlexNet algorithm with amyloid PET images. All subjects had an intravenous injection of 300 MBq of F-18-Florbetaben (F-18-FBB, Piramal Imaging, Berlin), and PET acquisition was started 90 min after the radio-tracer injection. The image data were classified into NC, MCI and AD based on the findings of the neurologist. We performed data augmentation using a method such as flip, rotation, and GAN, in addition to pre-processing and data augmentation for AlexNet, in order to supplement a number of deficient data. Artificial intelligence (AI) learning was simulated using the Mini-batch Stochastic Gradient Descent (MSGD) algorithm, and the learning rate was 5e-5, and the epoch was 100. Using a CNN and the ‘AlexNet’ architecture, we successfully classified F-18-FBB PET images of AD from NC where the accuracy of the test data on trained data reached 98.14%. In this work, the CNN, which is a deep learning neural network architecture, was used in order to distinguish AD and MCI from the NC. Accuracy was 98.33%, NC recall was 99.16%, MCI recall was 95.83% and AD recall was 98.16% after learning the data by using rotation and LR flip to increase the number of data. The accuracy was increased by 4.96%, NC recall was increased by 10.68%, MCI recall was decreased by 0.23%, and AD recall was increased by 9.65% after learning the data by using DCGAN to increase the number of data to 4020. Accuracy and recall were improved when data were learned through data augmentation by rotation and by left and right reversal. However, the effect of learning data by increasing the data by using up-down reversal and data augmentation through DCGAN was almost insignificant.