Search

The latest advancements in neuroimaging techniques have contributed to studying the disordered human brain, but still, there is a lack of disease-specific brain reference to localise brain seeds and standardise the comparison across studies. The current study first evaluates the inter-subject consistency across Attention-Deficit/Hyperactivity Disorder (ADHD) in a publicly accessible resting-state functional magnetic resonance imaging (RS-fMRI) ADHD-200 dataset by proposing a new voxels similarity index (VSI) that integrates both inter-atlases flexibility (F') from previous studies with proposed intersubjects stability (S') measure to improve outcomes. Secondly, the study employs a dual-layer clustering-based parcellation strategy inspired by the resultant improved Master Atlas networks to examine the ADHD cerebral cortex. The suggested approach to enhancing the ADHD cortex parcellation uses spectral clustering for global structure identification, followed by hierarchical clustering for local refinements and granularity specification. The final connectivitydriven brain reference achieved an average homogeneity of 0.63, and the enhanced Master Atlas with (S') achieved 0.35 homogeneity, surpassing the original Master Atlas with 0.27. This discovery implies that the final brain reference offers a more accurate and reliable framework to examine the connections and functions of the brain in individuals with ADHD. [ABSTRACT FROM AUTHOR]

Many deep learning-based recommender systems have been proposed recently. Where they involve all the users in datasets to build the latent representation of input data to be used later for predicting the missing rates. Despite the fact that, users have different interests, and these differences reduce the model prediction accuracy. This paper proposed a novel cluster-based denoising autoencoder model (cluster-based DAE) for rate prediction recommender systems. Instead of building a single model, it builds K models by using k-means algorithm to divide the users into groups based on their preferences. Each group trains a DAE model to generate recommendations for its members. The proposed method was trained and tested with MovieLens (100K, 1M, and 10M) datasets where 80% of the data are used for training and 20% for testing. The performance of the proposed method compared against other state-of-the-art methods that use deep learning to build rate prediction models. It outperformed the other compared methods in term of mean absolute error (with 12.9%, 14.7%, and 22.3%) and root mean squared error (with 24.2%, 18%, and 21.1%) using MovieLens 100K, 1M, and 10M datasets respectively. [ABSTRACT FROM AUTHOR]