University of Twente Student Theses

Login

Towards an automated analysis for the delineation of focal epilepsy : single pulse electrical stimulation and machine learning

d'Angremont, E. (2018) Towards an automated analysis for the delineation of focal epilepsy : single pulse electrical stimulation and machine learning.

[img] PDF
1MB
Abstract:In patients with drug-resistant focal epilepsy, surgery can be considered. The goal is to remove the epileptogenic tissue, while sparing the eloquent cortex. Prior to surgery, a prolonged electroencephalography (ECoG) recording can assist in the delineation of epileptogenic tissue and functionality of the surrounding cortex. During these recordings, single pulse electrical stimulation (SPES) of the intra-cranial electrodes is performed to evoke pathological responses from the epileptogenic tissue, which occur $>$100 ms after stimulation. These responses are called delayed responses (DRs). In the UMC Utrecht, they are visually analyzed by use of time-frequency (TF-SPES) images from approximately 2 sec. around stimulation. Each image is scored by two human observers on the presence of an evoked DR in three different frequency bands, namely spikes (10-80 Hz), ripples (80-250 Hz) and fast ripples (250-520 Hz). This visual analysis is very labor intensive. An additional problem is that DRs are occasionally observed as a physiological phenomenon. In the first part of this research, we trained a support vector machine (SVM) and a convolutional neural network (CNN) with the aim to automatically detect and classify the DRs in TF-SPES images. The training data consisted of 47197 images from 15 patients, with the consensus of two human observers as ground truth. The algorithms were tested on a total of 11394 images from 4 other patients. For the SVM, 9 features were defined and extracted from each image. The CNN used the whole image as an input. Classification was based on 5 different outputs. The SVM achieved a sensitivity of 0.88 and a precision of 0.65 for DRs on the test data. For the CNN this was 0.96 and 0.42, respectively. Both models seem to have overfit on the underrepresented classes. Finally, the models were applied to data of 4 additional patients for comparison with human observers. For both models, the agreement with human observers was comparable to the inter-rater agreement for the spike and ripple frequency bands. We conclude that both models can be applied for a more efficient analysis of SPES. At the second part of this research, we investigated the possibility of a CNN to find features that can distinguish between pathological and physiological DRs in TF-SPES images. The model was trained on 662 images and tested on 74 images, gathered from 8 different patients. All images contained DRs and were labeled as originating from either inside or outside the seizure onset zone (SOZ). The model achieved a sensitivity of 0.63 and a precision of 0.29 for DRs originating from the SOZ. These unsatisfactory results can be due to the low amount of data. Alternatively, it is suggested that the difference in pathological and physiological DRs cannot be found in TF-SPES images.
Item Type:Essay (Master)
Clients:
Unknown organization, Nederland
Faculty:TNW: Science and Technology
Subject:44 medicine, 50 technical science in general, 54 computer science
Programme:Technical Medicine MSc (60033)
Link to this item:http://purl.utwente.nl/essays/74818
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page