Theses supervised by Prof. Dr. Ahmet Ademoğlu

4 theses · Boğaziçi University

DoctorateOpen AccessEN

Nörogörüntülemede tensör analizi

Acquisition of large amounts of data in neuroimaging research requires development of new methods that can disentangle the underlying information and reveal the features related to cognitive processes. This thesis attempts to propose new methods that favor the multimodality and multidimensionality of the brain data. The main difficulty for the fusion of imaging modalities is the discrepancies in their spatial and temporal resolutions as well as the different physiological processes they reflect. This problem is addressed by decomposing the EEG and fMRI data cast as tensors on both common and discriminant subspaces and computing the common spatial profile from the data on the cortical surface. The Granger causality analysis of brain connectivity is reformulated on tensor space enabling incorporation of tools developed in that area of research. The first approach on this analysis facilitated tensor methods for sparse representation of the connectivity patterns whereas the second method resolved them as atomic structures. General theory and computationally efficient algorithms are presented. The techniques are illustrated on the simultaneous EEG/fMRI recordings for the fusion model and on the fast fMRI data for the connectivity analysis. The proposed approaches may have a wide application area ranging from the early diagnosis of neurological diseases to the brain-computer interface studies.

ElectroencephalographyMagnetic resonance imagingSpectral decomposition+2
Esin Karahan Şenvardar
Boğaziçi University · Biyomedikal Mühendislik Enstitüsü
2015
00
Master'sOpen AccessEN

Inter and intra-individual behavioral variability predicted by neural activity in the multiple demand network

A wide range of cognitive tasks consistently identify a Multiple Demand (MD) network in frontal and parietal brain regions. Its activity is closely linked to executive functions (EFs) such as attention, task switching, solving novel problems and manipulating information in the working memory. We here investigate the relation between MD neural activity and EFs using a large fMRI dataset (n=120). We examine this relation through two approaches (1) inter-individual variability, addressing several methodological challenges: We find that MD activity - which varies substantially across individuals, but is consistent within individuals across time - can explain a substantial proportion of variance in individual performance on a spatial working memory task such that individuals who find the task challenging, increase their MD activity substantially to improve their performance. This suggests that MD activity tightly reflects the executive demand of an individual. In the second approach we examine (2) trial-by-trial variability by employing three different models to fuse reaction time (RT) data with the BOLD time-series. We find that BOLD amplitude increases with longer RTs. This is consistent with the findings from the first approach showing increased MD activity for slower individuals. Together both findings support the view that within and between individual differences are manifested in the same brain regions. These results have implications for (1) understanding brain processes of EFs through ID studies (2) given that ID in EF are largely genetically determined, genetic variability can be linked to the neural activity of the MD network as an intermediate stage to link genetics with behavior (3) using ID in fMRI responses as clinical biomarkers.

Individual differenceMagnetic resonance imagingModulation response+3
Moataz Assem
Boğaziçi University · Biyomedikal Mühendislik Enstitüsü
2016
00
Master'sOpen AccessEN

Epileptic seizure prediction using machine learning and deep learning methods

Epilepsy is one of the most common neurological diseases in the world which negatively affects the daily life of a patient. Predicting epileptic seizures is of great importance for healthcare professionals and patients. The electroencephalography (EEG), which allows for registering brain activity with the help of electrodes placed on the scalp, is generally used to diagnose and monitor epilepsy. In this study, automatic seizure prediction was performed using CHB-MIT dataset which contains EEG data recorded at Boston Children's Hospital. Support Vector Machines (SVM), a common machine learning algorithm chosen as the primary method within this thesis's scope, and three different deep learning methods were compared. The first of these methods was long short term memory (LSTM) classifier with convolutional autoencoder which did not need any feature extraction. The second method used the spectrograms obtained by preprocessing the EEG data which were fed into a convolutional neural network (CNN) based classifier. The last method was based on converting the EEG data into three-dimensional images by applying source localization and performing classification with CNN. Among the methods used, the best result was obtained using source localization based CNN classification with 89.06% specificity, 92.58% sensitivity and 90.41% accuracy. Computational cost of three methods in terms of runtime efficiency were also compared, and it was observed that the SVM, which yielded the lowest classification performance with 74.07% accuracy, worked significantly faster than other methods.

Burak Gözütok
Boğaziçi University · Biyomedikal Mühendislik Enstitüsü
2021
00
Master'sOpen AccessEN

Motor imagery classification using multivariate empirical mode decomposition (MEMD) based functional connectivity features

Electroencephalogram (EEG) motor imagery signals are widely used for the implementation of brain-computer interfaces (BCI). Recently, functional connectivity measures have attracted attention as they can be used to capture statistical dependencies among EEG channels. However, functional connectivity during motor imagery tasks have not been fully explored. This study utilizes Instrinsic Mode Function (IMF) level phase locking value (PLV), coherence, and imaginary part of coherency in lefthand/right-hand motor imagery classification. EEG signals are decomposed into IMFs via noise-assisted multidimensional empirical mode decomposition (NA-MEMD), and connectivity metrics over the selected four channels are calculated for each trial as raw data and as a function of time and frequency. Resulting features are used to train multiple classifiers and their accuracy scores are analyzed. Best results are obtained using features derived from time-frequency functions of imaginary part of frequency where the overall accuracy score of 0.77 is achieved. This study shows that the change of connectivity throughout the duration of the task provides a more effective feature than connectivity calculated using each trial as raw data. Achieved accuracy scores are comparable to similar studies with the additional advantage of using few number of channels.

Brain-computer interfaceMachine learning
Fatih Ekrem Onat
Boğaziçi University · Biyomedikal Mühendislik Enstitüsü
2023
00

Other supervisors