HLGSNet: Hierarchical and Lightweight Graph Siamese Network with Triplet Loss for fMRI-based Classification of ADHD Ranjeet Ranjan Jha, Aditya Nigam, Arnav Bhavsar School of Computing and Electrical Engineering Indian Institute of Technology Mandi, India d16044@students.iitmandi.ac.in, (aditya, arnav)@iitmandi.ac.in Gaurav Jaswal Department of Electrical Engineering Indian Institute of Technology Delhi Delhi, India gauravjaswal@ee.iitd.ac.in Sudhir K Pathak Learning Research and Development Center University of Pittsburgh Pittsburgh, USA skpathak@pitt.edu Rathish Kumar Department of Mathematics and Statistics Indian Institute of Technology Kanpur Kanpur, India bvrk@iitk.ac.in Abstract—Attention Deficit Hyperactivity Disorder (ADHD) is a behavior-based disorder that mainly occurs in young children. Resting-state fMRI data have been very popular for diagnos- ing brain disorders like Autism, ADHD, and schizophrenia, by network-based functional connectivity, since these disorders are associated with both individual brain regions and their connections. Finding patterns among regions of controls’ brain and ADHD patients’ discriminating brains, is a non-trivial task. For classification of ADHD, we propose an end-to-end lightweight CNN architecture with hierarchical representation learning i.e., HLGSNet. We extract 116 anatomical regions from each subject in both normal and patient conditions, and graphs are built with the help of temporal correlation between different regions, where each region is considered as a node. Following this, a Siamese graph convolution neural network with triplet loss has been trained for finding embeddings so that samples for the same class should have similar embeddings. Finally, along with a fully connected layer, the trained model has been fine-tuned for the classification task. Experiments have been carried out on publicly available ADHD-200 dataset with promising performance. Index Terms—fMRI, Graph convolution neural networks, ADHD detection, Siamese Classification I. I NTRODUCTION One of the fastest-growing methods for assessing neural connectivity is the functional magnetic resonance imaging (fMRI). It is broadly categorized into two classes: task-based fMRI [1] and resting-state fMRI [2]. More recently, func- tional connectivity studies using rs-fMRI has been yielding promising detection results for various diseases, including Alzheimer’s, Attention deficit hyperactivity disorder (ADHD), Autism spectrum disorder (ASD), and epilepsy. ADHD is one of the most severe neuro-developmental diseases impacting 5-10% of young children between the ages of 6 and 12. Since no specific diagnostic approach is reported for ADHD, the diagnosis depends on findings, typically over months, by medical practitioners or parents. Thus, this is a very time consuming and costly process. However, non- invasive brain imaging techniques along with advanced signal processing methods can help in early detection of ADHD [3], [4]. Machine learning methods can play a crucial role in the discovery of the difference in brain connectivity patterns across ADHD and healthy subjects. Many approaches, such as clustering [5], sparse dictionary learning [6], correlation [7], and graph-based techniques [8], have been used for feature extraction and selection, followed by a classifier to predict the actual class label. However, few existing classification techniques use hand-crafted features extractor and traditional machine learning frameworks which do not have automatic feature learning capability. Recently, deep learning methods [9], [10] are being explored to classify subjects based on the functional connectivity of brain regions. Challenges and Contribution: There have been several scientific attempts to identify and recognize brain diseases, using fMRI activation. However, these methods are usually not generalizable to smaller data sets because traditional networks tend to get over-fitted and do not perform well on testing datasets. Moreover, conventional graph networks fail to learn hierarchical representations. To overcome aforementioned lim- itations, this paper presents an end-to-end CNN architecture, i.e., Hierarchical and Light-weight Graph Siamese Network (HLGSNet), for classification of neurological disorders, par- ticularly ADHD from fMRI data. Our contribution in this paper is of four folds: (1) Each region has been represented by average signal of that region which helps to get the denoised signal unlike at voxel level. (2) The network is trained in a Siamese framework using a triplet loss function where graph convolution features are learned from the graph 978-1-7281-6926-2/20/$31.00 ©2020 IEEE