EEG based multi-class seizure type classification using convolutional neural network and transfer learning.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate epileptic seizure type recognition is crucial for neurosurgical planning and understanding brain connectivity.
• Existing automated seizure detection methods lack classification for seizure variants.
• This study addresses the need for multi-class classification of diverse epileptic seizure types.

Purpose of the Study:

• To classify seven variants of epileptic seizures and non-seizure electroencephalogram (EEG) data.
• To apply convolutional neural networks (CNNs) and transfer learning for seizure classification.
• To evaluate the efficacy of different pre-trained CNN models for this task.

Main Methods:

• Utilized the Temple University Hospital EEG corpus, converting 19-channel EEG time series into spectrogram stacks.
• Employed two CNN-based modalities: direct transfer learning and feature extraction with a support vector machine classifier.
• Tested ten pre-trained networks including Alexnet, Vgg16, Vgg19, Squeezenet, Googlenet, Inceptionv3, Densenet201, Resnet18, Resnet50, and Resnet101.

Main Results:

• Achieved the highest classification accuracy of 82.85% with Googlenet using transfer learning.
• Obtained a superior accuracy of 88.30% with Inceptionv3 when extracting image features.
• Demonstrated that CNN-based approaches significantly outperformed conventional feature and clustering-based methods.

Conclusions:

• EEG-based seizure type classification using CNN models is a viable approach for pre-surgical evaluation in epilepsy patients.
• Transfer learning and feature extraction with CNNs offer high accuracy in differentiating seizure variants.
• This methodology holds promise for improving diagnostic accuracy and treatment planning in epilepsy management.