The expert's knowledge combined with AI outperforms AI alone in seizure onset zone localization using resting state...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Refractory epilepsy (RE) necessitates precise seizure onset zone (SOZ) localization for effective surgical treatment.
• Resting-state functional MRI (rs-fMRI) and deep learning (DL) show promise but require refinement for SOZ identification.

Purpose of the Study:

• To evaluate if integrating expert epileptologist guidance with DL techniques improves SOZ delineation in RE patients compared to DL alone.
• To explore the characteristics of expert-identified SOZ patterns.

Main Methods:

• rs-fMRI data from 52 children with RE were analyzed.
• Functional connectomics data were classified by experts and used to train an expert knowledge-integrated DL model.
• A DL-only model was also trained for comparison.

Main Results:

• The expert knowledge-integrated DL model achieved 84.8% accuracy and a 91.7% F1 score for SOZ localization.
• The DL-only model achieved less than 50% accuracy (63% F1 score).
• Discriminative SOZ characteristics involved gray matter activation extending through white matter to vascular regions.

Conclusions:

• Integrating expert knowledge with DL significantly enhances SOZ localization accuracy in RE patients.
• This hybrid approach offers potential explanations for SOZ co-activation patterns.
• Preoperative rs-fMRI studies combined with surgical outcomes can yield crucial expert knowledge for SOZ identification.