ADHD Classification Method based on Spatio-Temporal Multifractal Features Using Attention-Based Pyramid Feature Mixer Modules

Attention Deficit Hyperactivity Disorder (ADHD) is characterized by atypical functional brain connectivity, although its exact cause is not fully understood. To enhance diagnostic accuracy, researchers use resting state functional magnetic resonance imaging (rs-fMRI) to extract functional connectivity features. In this paper, we propose a novel ADHD diagnosis method with an attention-based pyramid model, which leverages spatio-temporal multifractal features from rs-fMRI to enhance the diagnosis performance. Firstly, we separate the rs-fMRI to 20 independent components (ICs), each associated with 3D spatial maps and corresponding 1D time-series. Next, we employ multifractal analysis to extract temporal self-similar features from each time-series. To further utilize the global information present in the ICs, we introduce an Attention-based Pyramid Feature Mixer (APFM) module. The 3D spatial features from each IC are processed through a 3D-APFM, while the temporal self-similar features are passed through a 1D-APFM. The outputs from both the 3D and 1D-APFM modules are fused to predict the final diagnosis results. We evaluate the performance of our proposed method and modules using 5 datasets from ADHD 200. Our method achieves an accuracy of 90.1% and sensitivity of 97.6%, both of which are substantially higher than the previous state-of-the-art performance. Therefore, we suggest that the spatio-temporal multifractal feature captures essential information during changes in brain activity, allowing it to describe underlying stochastic properties of ADHD rs-fMRI. At the result, when combined with APFM modules, these features significantly enhance the performance of ADHD diagnosis.