Planning stereoelectroencephalography using automated lesion detection: Retrospective feasibility study

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Drug-resistant epilepsy in pediatric patients presents complex diagnostic challenges.
• Stereoelectroencephalography (sEEG) is crucial for localizing the seizure onset zone (SOZ) in epilepsy surgery planning.
• Accurate identification of epileptogenic lesions on structural magnetic resonance imaging (MRI) is critical for successful sEEG implantation.

Purpose of the Study:

• To evaluate the feasibility and benefits of integrating deep learning-based MRI lesion detection into sEEG planning for pediatric drug-resistant epilepsy.
• To assess the colocalization between automated lesion detection and the SOZ identified by sEEG.

Main Methods:

• A neural network classifier was trained on MRI data from patients with focal cortical dysplasias (FCDs) and healthy controls.
• The classifier's performance was evaluated in a cohort of pediatric patients undergoing sEEG implantation.
• sEEG contact coordinates were coregistered with classifier-predicted lesions to determine colocalization with the SOZ.

Main Results:

• The deep learning classifier achieved 74% sensitivity in detecting radiologically defined lesions and 100% specificity in healthy controls.
• In sEEG patients, the algorithm correctly identified 86% of histopathologically confirmed FCDs.
• A high degree of colocalization (62%) was observed between classifier-detected lesions and the SOZ in patients with heterogeneous focal cortical lesions.

Conclusions:

• Deep learning-based automated MRI lesion detection shows high colocalization with the SOZ identified by sEEG in pediatric epilepsy.
• A framework for incorporating this technology into sEEG implantation planning has been developed.
• Prospective evaluation of automated MRI analysis is supported for optimizing electrode trajectories in epilepsy surgery.