Deep Learning–Based 3D Convolutional Neural Networks for ADHD Classification Achieving Over 60% Accuracy: A Comprehensive Study
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Attention-Deficit/Hyperactivity Disorder (ADHD) is a complex neurodevelopmental condition that affects millions of individuals worldwide, significantly impacting their academic performance, social interactions, and overall quality of life. Recent advances in neuroimaging techniques and machine learning methodologies have made it possible to analyze high-dimensional 3D brain images with increasing precision. In this study, we present a data-driven approach using 3D Convolutional Neural Networks (3D-CNNs) for the classification of ADHD versus non-ADHD (control) participants. Our dataset includes 691 phenotypic entries and 896 normalized, resampled 3D MRI scans derived from the ADHD-200 Preprocessed dataset. After merging phenotypic data with brain images based on matching subject IDs, we obtained 605 valid samples (423 for training, 91 for validation, and 91 for testing).We implemented two distinct 3D deep learning models: (1) a custom Simple3DCNN architecture, and (2) a 3D ResNet (r3d_18) from PyTorch’s video models. Both models were trained on CPU in a Kaggle-like environment for two epochs with a standard learning rate of 1e-4. Evaluation on the held-out test set yielded a classification accuracy of approximately 62.64% and 61.54% for Simple3DCNN and 3D ResNet, respectively. The Simple3DCNN achieved an AUC of 0.6832, while the 3D ResNet achieved an AUC of 0.5965. Although our current results do not surpass 70% accuracy, they demonstrate the feasibility of applying 3D deep learning methods to ADHD MRI classification tasks and highlight specific challenges, such as imbalanced labels and subtle neuropathological differences, that require further methodological refinement.Our extensive discussion includes data preprocessing, hyperparameter tuning challenges, confusion matrix analyses, and directions for future work. We also provide a thorough literature background and show that a single run with minimal epochs is unlikely to achieve near-perfect classification in such a heterogeneous dataset. Nevertheless, these findings serve as a promising step toward robust biomarkers for ADHD using advanced deep learning in neuroimaging research.