ADHDNet: A DNN Based Framework for Efficient ADHD Detection from fMRI Dataset

The study of functional connectivity is an evolving field of research in brain network-based analysis of neurological disorders. The interconnection between various brain regions is affected due to different neuro-disorders. Attention Deficit Hyperactivity Disorder (ADHD) has been studied using complex network-based features from the ADHD-200 competition fMRI dataset. The objective is to classify a typically developing subject from one showing a mature stage of acute symptoms (e.g., insufficient attention and/or hyperactivity). ADHD symptoms, being difficult to diagnose efficiently, will, if successfully detected computationally, lead to suitable clinical intervention and improved outcomes. The paper’s novelty is to capture the change of functional connectivity between brain regions of interest (ROIs) due to the ADHD syndrome using the brain atlas (MSDL, BASC-64/444) and complex network measures. A novel 5-layered Deep Neural Network (ADHDNet) has been implemented in this paper for efficient computer-aided diagnosis of ADHD. The output is compared with the traditional and best-performing machine learning technique Gradient Boosting as the dataset used is imbalanced between control population and mature-ADHD patients. SMOTE, Random Under (RUS), and Over (ROS) Samplers have been employed to deal with the data imbalance. This study is unique in its focus on the efficient detection of ADHD cases using complex network concepts as the feature extractor. The best performing results are 100% and 93% test-accuracies from BASC-64 + RUS and MSDL + ROS, respectively. The proposed ADHDNet provides consistently excellent and stable results based on evaluation metrics such as F1-score, accuracy, and Area under the ROC Curve(AUC).