Significant CC400 Functional Brain Parcellations Based LeNet5 Convolutional Neural Network for Autism Spectrum Disorder Detection

Abstract

Machine learning and computer vision have opened new pathways to investigate imaging data captured from different sensors. Numerous application areas are getting benefit from these advancements and one of these areas is medical imaging. Despite rapid advancements in machine learning based medical condition diagnosis systems (CADs), some ailments and disorders are hard to diagnose/classify due to the absence or the lack of consensus on biomarkers for specific disorders, like the Autism Spectrum Disorder (ASD). In this study, the challenging problem of classification of ASD using the magnetic resonance imaging (MRI) data is tackled. Hence, we propose an interpretable deep neural network based approach for ASD detection from MRI images. Our proposed explanation method is based on the selection of four regions of interest from the MRI images. The four significant CC400 functional brain parcellations are then concatenated and fed to a LeNet-5-based convolutional neural network to predict ASD. The performances of the proposed approach are evaluated on ABIDE dataset and promising results are achieved. Three augmented datasets are considered and an accuracy of 95 % is achieved by using LeNet-5 which outperforms VGG16 and ResNet-50. The achieved accuracy outperforms also the existing deep neural networks based approaches on ABIDE dataset. The use of the four significant CC400 functional brain parcellations makes our approach more interpretable and more accurate.