Identifying Epilepsy Based on Deep Learning Using DKI Images.

Area of Science:

• Neuroimaging
• Medical Diagnostics
• Epilepsy Research

Background:

• Refractory epilepsy is challenging to treat, often requiring surgery.
• Accurate lesion localization is difficult, especially in MRI-negative cases.
• Identifying epileptic foci aids in surgical planning and treatment.

Purpose of the Study:

• To investigate the utility of diffusion kurtosis imaging (DKI) parameters for epilepsy detection in children.
• To develop a classification model for distinguishing epilepsy patients from healthy controls using hippocampus-based DKI features.
• To evaluate the effectiveness of a convolutional neural network (CNN) for feature extraction and classification.

Main Methods:

• Recruited 59 children with epilepsy and 70 controls.
• Acquired diffusion kurtosis images (DKI) for all subjects.
• Segmented hippocampus and extracted fractional anisotropy (FA), mean diffusivity (MD), and mean kurtosis (MK) parameters.
• Utilized a CNN for feature extraction and a support vector machine (SVM) for classification.

Main Results:

• The CNN-based classifier achieved 90.8% accuracy in distinguishing epilepsy patients from normal controls.
• Fractional anisotropy (FA) and mean kurtosis (MK) were identified as significant features for epilepsy identification.
• CNN-based feature extraction proved effective for this classification task.

Conclusions:

• DKI parameters, particularly FA and MK, can serve as valuable biomarkers for epilepsy diagnosis.
• CNNs offer an effective approach for analyzing DKI data and classifying epilepsy.
• This study highlights the potential of DKI in improving clinical diagnosis of epilepsy, especially in challenging cases.