Deep Learning-Based Visual Complexity Analysis of Electroencephalography Time-Frequency Images: Can It Localize the...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Drug-resistant epilepsy necessitates precise localization of the epileptogenic zone (EZ) to guide neurosurgery.
• Visual inspection of intracranial electroencephalography (iEEG) signals and time-frequency (TF) images is standard but may miss subtle indicators.

Purpose of the Study:

• To develop and evaluate a deep learning-based metric of visual complexity for interpreting iEEG TF images.
• To assess the metric's ability to identify the EZ in patients with drug-resistant epilepsy.

Main Methods:

• Analyzed iEEG data from 20 children with drug-resistant epilepsy using 1928 contacts.
• Generated TF images (1-70 Hz) and measured visual complexity using unsupervised activation energy (UAE) from a pre-trained VGG16 network.
• Identified potential EZ using UAE values and a support vector machine classifier.

Main Results:

• Contacts within the seizure onset zone showed significantly lower UAE compared to those outside, particularly in deeper convolutional layers (p < 0.001).
• The support vector machine approach localized the EZ with an accuracy of 7 mm.

Conclusions:

• A novel deep learning tool utilizing UAE offers a computerized approach for EZ localization.
• This method aids pre-surgical planning by interpreting iEEG TF images, reducing reliance on extensive visual inspection.