Leveraging Models of Human Reasoning to Identify EEG Electrodes in Images With Neural Networks

Leveraging Models of Human Reasoning to Identify EEG Electrodes in Images With Neural Networks

Alark Joshi, Phan Luu, Don M. Tucker, Steven Duane Shofner
Copyright: © 2019 |Pages: 28
DOI: 10.4018/978-1-5225-5751-7.ch005
OnDemand:
(Individual Chapters)
Available
$37.50
Current Special Offers
No Current Special Offers
TOTAL SAVINGS: $37.50

Abstract

Humans have very little trouble recognizing discrete objects within a scene, but performing the same tasks using classical computer vision techniques can be counterintuitive. Humans, equipped with a visual cortex, perform much of this work below the level of consciousness, and by the time a human is conscious of a visual stimulus, the signal has already been processed by lower order brain regions and segmented into semantic regions. Convolutional neural networks are modeled loosely on the structure of the human visual cortex and when trained with data produced by human actors are capable of emulating its performance. By black-boxing the low-level image analysis tasks in this way, the authors model solutions to problems in terms of the workflows of expert human operators, leveraging both the work performed pre-consciously and the higher-order algorithmic solutions employed to solve problems.
Chapter Preview
Top

Background

Philips Neuro manufactures complete dense array electroencephalography (EEG) systems. EEG, as a technology, has existed since 1875 (Swartz, 1998) and is much older than many other technologies used for non-invasive neural imaging, such as magnetic resonance imaging (MRI) or functional magnetic resonance imaging (fMRI), but offers some unique advantages such as high temporal resolution and movement tolerance.

Epileptic seizures result from excessive synchronization in the firing of neurons in the brain (often starting in a small highly localized region). The high temporal resolution (often 1,000 samples per second or more) makes EEG ideal for the diagnosis of epilepsy, as the associated spikes are most easily characterized by the millisecond resolution enabled by EEG. In some cases, epileptic seizures are not controlled by pharmacological means and a determination is made that a brain resection should be performed to remove the tissue that serves as the source of the problem. When resectioning is deemed appropriate source localization of the recovered electrical signals can assist in identifying the region of the brain responsible. In order to perform source localization using the recovered electrical signals, the sensor locations must be registered to the surface of a model of the head.

There are a number of other situations where the ability to localize recovered signals is highly desirable. There is some evidence that neuronal activation can be depressed by low frequency pulsed direct current stimulation (Groppa et al., 2010). This may offer a non-invasive means for the treatment of neurological disorders like Epilepsy or Parkinson’s. For these disorders, it may be possible to use source localization of signals in EEG to identify the region of origin of the unwanted activity and then, given the reciprocity principle (Fernández-Corazza, Turovets, Luu, Anderson, & Tucker, 2016) (which holds that electrical current will follow the same pathways when injected into the scalp as it follows to reach the scalp from a specific brain region), target the same region with transcranial injected current to modulate the activation or plasticity of the region.

In both cases, an accurate registration of the sensors to the scalp surface is needed. Several means of achieving this exist with their own advantages and disadvantages.

Complete Chapter List

Search this Book:
Reset