DeepEZ: A Graph Convolutional Network for Automated Epileptogenic Zone Localization From Resting-State fMRI...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate epileptogenic zone (EZ) localization is critical for managing medication-refractory epilepsy.
• Traditional methods often focus on localized signatures, but seizures impact broader brain connectivity.

Purpose of the Study:

• To present the first deep learning approach, DeepEZ, for EZ localization using resting-state fMRI (rs-fMRI).
• To develop a noninvasive tool to aid in therapeutic planning for epilepsy patients.

Main Methods:

• DeepEZ employs a cascade of graph convolutions to model signal propagation along anatomical pathways.
• The model integrates domain-specific information, including an asymmetry term and subject-specific bias, to account for environmental confounds.

Main Results:

• DeepEZ demonstrated high performance in EZ localization on rs-fMRI data from 14 focal epilepsy patients (Accuracy: 0.88 ± 0.03; AUC: 0.73 ± 0.03).
• Performance remained robust despite variability in EZ locations and scanner types across the cohort, significantly outperforming baseline methods.

Conclusions:

• DeepEZ shows significant promise as an accurate and noninvasive tool for therapeutic planning in medication-refractory epilepsy.
• This deep learning approach effectively leverages inter-regional connectivity information from rs-fMRI, integrating seamlessly into clinical workflows.