Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterized by persistent patterns of inattention, hyperactivity, and impulsivity. With the recent advances in neuroimaging and machine learning models, automatic ADHD diagnosis has been made possible. Most of the studies on ADHD detection focus on the use of functional connectivity patterns. This chapter proposes a novel approach to enhance ADHD detection using functional connectivity analysis with two main contributions. Firstly, an autoencoder is employed as a feature selector to effectively reduce dimensionality and extract discriminative features from functional connectivity matrices. Thanks to this approach, the dimensionality of the feature space is reduced and the robustness of the classification model is increased. Secondly, instead of considering all brain regions, this study focuses on regions of interest (ROIs) within the default mode network (DMN), a network implicated in ADHD pathology. The model performance in identifying ADHD-related connectivity patterns is explored by concentrating on different DMN regions proposed in literature. ADHD-200 database is used as a benchmark for comparison with previous studies demonstrating promising results in terms of classification accuracy.