Electrocorticographic Features and Classification of Intracranial Glioma

Abstract

During surgical resection of tumorous intracranial tissue in the human brain, different methods are used to locate the boundaries of the tumorous tissue because maximal resection of the tumorous tissue has been associated with better post-surgical outcomes for patients [1]. Electrocorticography (ECoG) is one method used intraoperatively to classify tissue as tumorous or healthy tissue. The electrophysiological features of neural tissue in the presence of tumorous intracranial tissue remain relatively unexplored. The exploration of these electrophysiological features may lead to more accurately defined tumorous tissue boundaries and improved post-operative outcomes. In this thesis study the Power Law Exponent (PLE) is explored as a feature for use in the classification of tumorous tissue. Spectral amplitudes and functional connectivity features are also explored as features for use in the classification of tumorous tissue. Three different classifier types, discriminant analysis, support vector machines (SVMs), and artificial neural networks (ANNs), are used with the features explored in this paper to study the relative values of the features for use in the generalized classification of tumorous tissue. ECoG datasets from 17 subjects were selected from a pool of 23 awake, resting state datasets belonging to subjects with WHO grade II-IV intracranial glioma. ECoG electrodes labelled “healthy” or “tumorous” were studied using PLE, spectral amplitude, and functional connectivity features and it was found that the PLE feature was the most valuable in the classification of tumorous tissue. The largest significant difference between the mean values of the distributions of healthy and tumorous tissue was found in the case of the PLE feature (𝜇 - ℎ𝑒𝑎𝑙𝑡ℎ𝑦 = 3.676, 𝜎^2 - ℎ𝑒𝑎𝑙𝑡ℎ𝑦 = 0.178, 𝜇 - 𝑡𝑢𝑚𝑜𝑟𝑜𝑢𝑠 = 3.214, 𝜎^2 - 𝑡𝑢𝑚𝑜𝑟𝑜𝑢𝑠, 𝑝 = 6.224x10^−20). A three-layer MultiLayer Perceptron (MLP) ANN was firstly trained on all features and secondly trained on all features without the PLE. The model trained with all features achieved a correct classification rate of 79.3% while the model trained without PLE achieved a correct classification rate of 62.7%.