Multicenter Validation of a Deep Learning Detection Algorithm for Focal Cortical Dysplasia.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Epilepsy Research

Background:

• Focal cortical dysplasia (FCD) is a leading cause of drug-resistant epilepsy.
• Conventional MRI often fails to detect FCD, leading to 'MRI-negative' epilepsy diagnoses.
• Accurate FCD detection is crucial for effective surgical planning and patient outcomes.

Purpose of the Study:

• To develop and validate a multicenter deep learning algorithm for detecting FCD on MRI.
• To assess the algorithm's performance, particularly in MRI-negative cases.
• To evaluate the algorithm's generalizability across different centers and MRI hardware.

Main Methods:

• A deep convolutional neural network (CNN) was trained using 3D T1-weighted and 3D fluid-attenuated inversion recovery MRI data from 148 patients with histologically verified FCD across 9 centers.
• Bayesian uncertainty estimation was incorporated for risk stratification.
• Performance was evaluated by comparing algorithm detection maps to expert labels and tested on an independent cohort of 23 FCD cases and 131 controls (42 healthy, 89 with temporal lobe epilepsy).

Main Results:

• The algorithm achieved an overall sensitivity of 93% and 85% for MRI-negative FCD.
• In an independent cohort, sensitivity was 83%.
• Specificity was 89% in healthy and disease controls, with an average of 5-6 false positives per patient.

Conclusions:

• This multicenter-validated deep learning algorithm demonstrates high sensitivity for detecting FCD, especially in MRI-negative cases.
• The algorithm's ability to provide risk stratification enhances diagnostic confidence.
• Its generalizability makes it a valuable tool for presurgical evaluation in epilepsy management.