A multimodal deep learning framework for functional brain network classification in rs-fMRI.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Epilepsy diagnosis and surgical planning rely on accurate identification of functional brain networks.
• Resting-state functional magnetic resonance imaging (rs-fMRI) provides valuable data for analyzing brain connectivity.
• Automating the classification of rs-fMRI-derived networks can reduce subjectivity and workload in clinical practice.

Purpose of the Study:

• To develop and validate a deep learning framework for automated classification of functional brain networks in epilepsy patients using rs-fMRI.
• To differentiate between seizure onset zones (SoZ), resting-state networks (RSNs), and artifact/noise components.
• To assess the contribution of spatial, temporal, and spectral features in network classification.

Main Methods:

• A hybrid deep learning architecture combining a 3D Convolutional Neural Network (3D-CNN) for spatial features (SF) and a Long Short-Term Memory (LSTM) network for temporal (TS) and frequency-domain (FS) signals was employed.
• Independent Component Analysis (ICA) was used to derive functional brain networks from rs-fMRI data.
• An ablation study evaluated the impact of different feature types (SF, TS, FS) on classification performance.
• Expert neurologists provided qualitative validation of the model's interpretability and clinical relevance.

Main Results:

• The hybrid model achieved up to 70% accuracy in classifying 11 distinct ICA component types, including SoZ.
• Incorporating frequency-domain signals (SF+FS) improved ROC AUC to 0.78, while combining all features (SF+TS+FS) yielded the highest accuracy.
• The 'Noise' class showed high performance (up to 0.94), whereas the 'Temporal' lobe network class had lower scores (0.14-0.24).

Conclusions:

• The developed data-driven deep learning framework effectively automates the classification of rs-fMRI-derived functional brain networks in epilepsy.
• The integration of spatial, temporal, and spectral features enhances classification accuracy and supports clinical applications like epilepsy surgical planning.
• This approach has the potential to reduce subjectivity and improve efficiency in the clinical review of brain imaging data for epilepsy patients.