A HighPerformance Method Based On Features Fusion of EEG Brain Signal and MRIImaging Data For Epilepsy Classification
A HighPerformance Method Based On Features Fusion of EEG Brain Signal and MRIImaging Data For Epilepsy Classification
1, 1-8
Abstract: A 1-dimensional (1D) and 2-dimensional (2D) biomedical signal analysis based on the Discrete Cosine Transform (DCT) feature
extraction method was performed to diagnose epilepsy disorders with high accuracy. For this purpose, Electroencephalogram (EEG) data
were used for 1D signal analysis and Magnetic Resonance Imaging (MRI) data were used for 2D signal analysis. The feature vectors were
obtained by applying 1D DCT together with statistical methods such as mean, variance, standard deviation, kurtosis, and skewness for EEG
data and by applying 2D DCT together with the statistical method of mean for MRI data. The most useful features were selected by applying
Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Forward Selection and Backward Selection methods to the
obtained feature vectors. Using EEG stand-alone features, MRI stand-alone features and EEG-MRI fused features, the classification of
healthy and epileptic subjects was performed in the form of two clusters. The result of epilepsy classification in this work is 96% success of
1D EEG data by using the features selected by the PCA method, 94% success of 2D MRI data using the selected features by applying the
Forward Method, 100% classification accuracy of 1D EEG and 2D MRI datasets by LDA method using the obtained fused features . The
article shows that the fused features of EEG-MRI can be used very effectively for the diagnosis of epilepsy.
Keywords: Epilepsy, MRI, EEG, discrete cosine transform, medical image analysis, EEG-MRI classification.
1
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
2. EEG ANALYSIS
The EEG is a test in which the electrical activity in the
brain is measured via electrodes attached to the scalp. The
EEG is used to diagnose certain diseases that affect the brain. Fig. 2. Nihon Kohden tool.
Epilepsy is the most common one. There are many steps from
D. Pre-processing
obtaining the EEG data to classifying it. Classification is
about whether the person has epilepsy based on the EEG data. The EEGLab14_1_2b program was used to examine and
pre-process the EEG data and remove artifacts. The data then
A. Materials went through many stages such as feature extraction,
selection and classification.
The Neurofax EEG-1000 was used to record the EEGs of
the patients diagnosed with epilepsy in the Neurology 1. Data reading
Department of Bolu Abant-Izzet Baysal University Hospital The plugin 1.1.0 from Nihon Kohden is added to EEGLab
in Bolu. In total, the EEG data of 88 epilepsy patients and 40 to solve this problem (due to the unsupported data format).
healthy individuals were recorded. The EEG cap has 16 The EEG data (.m00 format) was transferred to the system
channels. The recording takes an average of 18 minutes for via the EEGLab installed plugin. The 16-channel EEG signal
each patient. The computer system used for the application is received in the system was terminated in 1080 seconds with
a transfer rate of 200 Hz in a matrix of 16x21600, and the size
an Intel (R) Core (TM) i5-4210U CPU 1.70 GHz, x64 based
of the data set was determined to be 15.6 Mb.
processor.
2. Filtering
B. Method
When raw data is read, this data also contains unwanted
The EEG data is read out with the Nihon Kohden EEG- noise. A high-pass filter with a value between 0.5 – 1 Hz was
1000 fax program, the software of the EEG device. Fig. 1 applied. In the study, the unwanted noise contained the raw
shows the stages of the EEG from the analysis of the first raw data is removed by a high-pass filter when the data is read.
data input to the classification step. The noise is removed from the data using the filter.
2
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
3. Channel cleaning 2
𝑢 ≠ 0 𝑤ℎ𝑒𝑛 𝐶(𝑢) = √
As the filtered data was limited to the frequency values 𝑁
3
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
4
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
3. 2D DCT mean
This method consists of using the aforementioned two
methods as a hybrid. The size of the obtained DCT
coefficients is 240x240 pixels. The pixels are divided
vertically and horizontally into eight equal, small parts. This
results in 64 subsections. The size of each subsection is 30x30
pixels. The arithmetic mean is calculated for each subsection.
Thus, the size of the new feature vector is 64 as described by
Demirezen-Yagmur and Sertbas [24].
G. Feature selection
The feature vector extracted using the DCT_mean method
is 64. The PCA, LDA, Forward Selection and Backward
Fig. 5. Segmentation [24].
Selection algorithms were applied to select the most
E. Pre-processing compelling feature.
Certain pre-processing was performed on the segmented H. Classification
MRI data. The matrices obtained as a result of the
The classification settings for EEG data also apply to MRI
segmentation are transformed to 240x240 and resized.
data. The data was trained separately according to the results
Dimensional MRI matrices are values ranging from 0 to 999,
of the different feature selections and the results are shown in
which were then converted to normalized values between 0
Table 2.
and 1.
Table 2. Different classification algorithm validation results
F. Feature extraction according to the feature selection methods of the MRI data.
In pattern recognition, feature extraction methods are
Method ANN SVM KNN
needed to reduce the size of the pattern. These methods are
mathematical functions that perform feature extraction and PCA 90% 81% 81%
selection. New features were extracted from MRI images by LDA 81% 81% 82%
the hybrid use of DCT and Arithmetic Mean methods, the Forward Selection 81% 81% 83%
transformed structural feature extraction methods known in
Backward Selection 81% 81% 81%
the literature, Meyer-Bäse [25].
1. 2D Discrete Cosine Transform (DCT) The success of the feature vectors selected by the PCA
method in classification is therefore high. The added metric
DCT is particularly useful for easy and fast feature
was needed because other recall and precision metric values
extraction from large data sets. It also plays a vital role in
were very close to each other. If such metric values are
image compression. After pre-processing, segmentation and
critical, the F1-score metric result is also examined. The PCA
normalization of the MRI data, the 2D (3) was applied and
method Backward Selection is shown in Table 5.
feature extraction was performed. In (8), u, v is the value of
the MRI matrix and N, M is the length of the matrix. 4. EEG-MRI FEATURES FUSION APPROACH
𝑁−1 𝑀−1
𝜋(2𝑥 + 1)𝑢 𝜋(2𝑦 + 1)𝑣 A new set of features is created by combining biomedical
𝐹(𝑢, 𝑣) = 𝐶(𝑢)𝐶(𝑣) ∑ ∑ 𝑓(𝑥, 𝑦) cos [ ] cos [ ] signals of different sizes. This feature vector consists of a
2𝑁 2𝑁
𝑥=0 𝑦=0
combination of 1D EEG and 2D MRI features of one and the
𝑈𝑝 𝑡𝑜 𝑢 = 0,1, … 𝑁 − 1 𝑣 = 0,1, … 𝑀 − 1 (8) same person.
A. Material
1 After applying the 1D DCT with Mean method to EEG
𝑢, 𝑣 = 0 𝑤ℎ𝑒𝑛 𝐶(𝑢), 𝐶(𝑣) = √
𝑁 data, 64 features were obtained. Similarly, 64 features were
obtained after applying the Mean method with 2D DCT on
MRI data. In the application, 1D and 2D biomedical signals
2
𝑢, 𝑣 ≠ 0 𝑤ℎ𝑒𝑛 𝐶(𝑢), 𝐶(𝑣) = √ of different sizes are fused, and a total of 128 feature data are
𝑁 obtained. A vector of 128 data features contains both EEG
and MRI features. In our study, a 128-feature data set was
2. Arithmetic mean obtained form 128 samples.
The mean is calculated by dividing the sum of all elements
in the series by the number of elements in the series. The B. Method
arithmetic mean is also referred to as the "mean" by The combination of EEG and MRI data, the feature
Abramowitz and Stegun [26]. selection and the classification flowchart are shown in Fig. 6.
5
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
6
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
7
MEASUREMENT SCIENCE REVIEW, 24, (2024), No. 1, 1-8
[22] Kumar, Y., Dewal, M. L., Anand, R. S. (2014). [26] Abramowitz, M., Stegun, I. A. (1972). Handbook of
Epileptic seizures detection in EEG using DWT-based Mathematical Functions with Formulas, Graphs, and
ApEn and artificial neural network. Signal, Image and Mathematical Tables; 9th printing. Dover Publications,
Video Processing, 8 (7), 1323-1334. ISBN 0-486-61272-4.
https://doi.org/10.1007/s11760-012-0362-9 [27] Gupta, A., Singh, P., Karlekar, M. (2018). A novel
[23] Gumaste, P. P., Jadhav, D. V. (2015). Image signal modeling approach for classification of seizure
segmentation techniques for brain MRI images: A and seizure-free EEG signals. IEEE Transactions on
survey. International Journal of Modern Trends in Neural Systems and Rehabilitation Engineering, 26 (5),
Engineering and Research (IJMTER), 2 (7). 925-935.
[24] Demirezen Yağmur, F., Sertbaş, A. (2018). Diagnosis https://doi.org/10.1109/TNSRE.2018.2818123
of epilepsy in brain MR images using DCT-means. In [28] Deivasigamani, S., Senthilpari, C., Yong, W. H. (2021).
2018 26th Signal Processing and Communications Machine learning method based detection and diagnosis
Applications Conference (SIU). IEEE, 2439-2443. for epilepsy in EEG signal. Journal of Ambient
https://doi.org/10.1109/SIU.2018.8404758 Intelligence and Humanized Computing, 12, 4215-
[25] Meyer-Bäse, A. (2004). Pattern Recognition in Medical 4221. https://doi.org/10.1007/s12652-020-01816-3
Imaging. Academic Press, 22-35.
https://doi.org/10.1016/B978-0-12-493290-6.X5000-7
Received May 10, 2023
Accepted January 09, 2024