Luigi Laura

Aim: Up to date many successful attempts to identify various types of lesions with machine learning (ML) were made, however, the recognition of Alzheimer’s disease (AD) from brain images and interpretation of the models is still a topic for the research. Here, using AD Imaging Initiative (ADNI) structural magnetic resonance imaging (MRI) brain images, the scope of this work was to find an optimal artificial neural network architecture for multiclass classification in AD, circumventing the dozens of images pre-processing steps and avoiding to increase the computational complexity. Methods: For this analysis, two supervised deep neural network (DNN) models were used, a three-dimensional 16-layer visual geometry-group (3D-VGG-16) standard convolutional network (CNN) and a three-dimensional residual network (ResNet3D) on the T1-weighted, 1.5 T ADNI MRI brain images that were divided into three groups: cognitively normal (CN), mild cognitive impairment (MCI), and AD. The minimal pre-processing procedure of the images was applied before training the two networks. Results: Results achieved suggest, that the network ResNet3D has a better performance in class prediction, which is higher than 90% in training set accuracy and arrives to 85% in validation set accuracy. ResNet3D also showed requiring less computational power than the 3D-VGG-16 network. The emphasis is also given to the fact that this result was achieved from raw images, applying minimal image preparation for the network. Conclusions: In this work, it has been shown that ResNet3D might have superiority over the other CNN models in the ability to classify high-complexity images. The prospective stands in doing a step further in creating an expert system based on residual DNNs for better brain image classification performance in AD detection.