Doç. Dr. Sami Arıca danışmanlığındaki tezler
11 tez · Çukurova University
A comparative study of distance/dissimilarity measures used for classification of electroencephalogram signals
Biological signal is a general term that refers to the signal measured from a biological system. Some representative examples include the electrocardiogram, the blood pressure waveform, the cellular action potential, etc. Many biological signals show distinctive waveform morphology which reflects the dynamics of the biological systems. The electroencephalogram (EEG) signal is a measure of the summed activity of approximately 1–100 million neurons lying in the vicinity of the recording electrode, and may provide insight into the functional structure and dynamics of the brain. Therefore, the exploration of hidden dynamical structures within EEG signals is of both basic and clinical interests. In clinical practices EEG is used to diagnose or monitor the following health conditions. In a computer aided diagnosis abnormal activities are detected and normal and abnormal activities are distinguished. An electroencephalograph (EEG)-based communication system, also known as brain–computer interface (BCI), utilizes the information in EEG and provide a new communication channel for patients with several motor disabilities, such as brain stem infarct or amyotrophic lateral sclerosis. The BCI requires classification or distinction of the information in EEG or state of EEG. All these applications necessitate distinction of state of the EEG. The distance or dissimilarity measures separability of the states of the EEG and provides that two or more EEG patters are different and each of them corresponds to a distinct state. In this study distance/dissimilarity measures for classifying EEG will be investigated and distinct features they recognize will be determined
Analysis and classification of evoked potentials of familiar and unfamiliar face stimuli
Electroencephalogram (EEG) is an electrical brain activity recorded from the electrodes placed on the skull surface. Evoked potential (EP) is an activity emerging in response to a specific stimulus in EEG. These potentials recorded from the brain are not only used in diagnosis of some diseases but also have recently been used in the design of applications enabling brain's interaction with computers and the environment. Examples of the applications are artificial limbs design, brain-computer design, oddball paradigm design and cerebral forensic inquiry design applications. In this study, familiar and unfamiliar face stimuli images were shown to different participants in different sessions during EEG recording and the acquired EEG data was analyzed and classified utilizing different methods. Analysis process comprised following stages respectively: pre-processing, feature extraction, channel selection, and classification and different methods were tried in each stage. At the pre-processing stage a low pass filter in time domain was used to suppress background signal and discrete cosine transform was employed to reduce spatial correlation between channels. The best classification performance was obtained when piecewise constant modeling was used at the feature extraction stage, mutual information was used at the channel selection stage and support vector machine was used at the classification stage. At the end of classification, brain electric activities were categorized in response to familiar and unfamiliar face stimuli. Consequently, this study will help us both in the pre-diagnosis of diseases such as Prosopagnosia and Alzheimer and in determining suspects' behavior in response to a face stimulus related to a crime instance in forensic inquiries.
Automatic segmentation of computed tomography images of liver using watershed and thresholding algorithms
Computed tomography (CT) imaging is widely used for control and diagnosis of diseases today. Segmentation of medical images is quite important especially for diagnosis and treatment of cancer. In this study, segments in CT images of liver are determined by using two different methods; watershed algorithm and histogram thresholding method. The images have been preprocessed before segmentation. First, images are converted to grayscale. Next, they are smoothed with a bilateral filter. To apply the watershed technique, edges are extracted with a Gradient operator. The over segmentation of the watershed method is overcome by merging the closest segments in terms of their features. The merging is obtained via vector quantization of the features; fuzzy c-means clustering and k-means clustering algorithms by grouping mean, standard deviation and features of segments which are also used in classification. The images are divided into five segments corresponding to liver, vertebra, tumour, lining and others. In case of histogram thresholding, multi thresholds are obtained with Otsu method from the smoothed image and segmentation has been performed. The results of two approaches have been compared. Pixel value, directional derivatives, local binary patterns, difference of pixel with its neighborhood haven been employed as features to determine the segment class. Classifications of the regions have been obtained from a single pixel with linear discriminant analysis classifier and segment with k- nearest neighbor classifier by dividing 25 liver images to two training (13 images) and test sets (12 images). The best accuracy was obtained % 95.57 for classification from a pixel with difference of pixel with its neighborhood feature whereas % 100 is obtained for categorization from the segment with directional derivatives and difference of pixel with its neighborhood features by using histogram thresholding algorithm. This application may help physicians in terms of providing insight with the tissues before the surgery by segmenting tissues in the medical images.
A simple approach to detect alcoholics using electroencephalographic signals
Electroencephalography (EEG) is a medical imaging technique that reads electrical activities generated by brain. In this study, electroencephalographic (EEG) signals acquired from alcoholics and controls have been analyzed. Raw EEG signals have been filtered with an 8-30 Hz bandpass filter. Normalization of EEG trials to a range [-1, 1] was performed to filtered EEG signal. We employed relative entropy and mutual information to specify the eight channel pairs with highest relative entropies and lowest mutual informations in rank from standart 10-20 electrode system (19 channels). Five channels which gives higher accuracy for classification of alcoholics and controls have been extracted. Feature vectors of training and test data were obtained by concatenating variances of these five channels. When relative entropy was used for channel selection, 80.33% accuracy was obtained with k-nearest neighbors classifier accompanied with Mahalanobis distance metric. And mutual information for channel selection process provided 82.33% accuracy with k-nearest neighbors classifier accompanied with Euclidean distance metric. Skewness and kurtosis probability distribution measures were used to extract training and test feature vectors and their classification performances were evaluated. At the end, to compare our results with a known feature extraction method, Common Spatial Patterns (CSP) algorithm were applied to the EEG data set to extract feature and results were compared with the methods used in this study. The results of the experimental analysis have been found satisfactory for alcoholic detection and may be useful in studying genetic predisposition to alcoholism.
Correlation between vector autoregressive model coefficients of circulatory system and baroreflex sensitivity
In this study, heart rate variability (RR) and systolic blood pressure (SBP) signals are considered as a two-channel signal. The two-channel signal has been represented by a first-order vector autoregressive model (VAR(1)) and autoregressive exogenous model (ARX), and baroreflex sensitivity (BRS) has been computed. This processing has been repeated for 17 subjects whose blood pressure has been altered with medicine (phenylephrine) injection and BRS, and VAR(1) model parameters have been related by employing multi-variable linear regression. The correlation coefficient between predicted BRS in this way and computed BRS has been obtained as 0.73. When linear relationship between BRS and VAR(1) model coefficients were examined, the correlation coefficient between BRS and b coefficient was found as 0.8132. The same study was repeated using ARX model and the correlation coefficient between BRS was found to be 0.72. The correlation coefficient between BRS and a12 coefficient of ARX model was found as 0.8096. These results shows that BRS can be predicted from VAR(1) and ARX model coefficients, and the used models characterize the circulatory system.
Biomedical image decomposition and segmentation by using signal expansion methods
Signal processing leads to research on many subjects. One of them is the biomedical field. Biomedical signal processing has played an important role in the detection process of many diseases in recent years. Different methods are used for purposes. Segmentation, separation, filtering, analysis are the leading methods. In this thesis, various signal expansion methods have been used for the decomposition and segmentation of biomedical images. Wavelet and adaptive filter banks-based methods have been used on medical images. Filter design with different optimization methods. These designed filters have been used to decompose and reconstruct images. The performances of the designed methods and classical methods were examined by comparing.
Detection of exudates from digital fundus images of diabetic retinopathy patients
Diabetes is a condition where the body does not produce enough insulin to convert sugar to energy, leading to a build up of sugar in the blood. This leads to a number of problems, including diabetic retinopathy. Diabetic retinopathy is a complication of diabetes that causes damage to the blood vessels of the retina that allowing you to see fine detail. It causes progressive damage to the retina. One of the earliest and most common symptoms of exudate diseases leading to blindness such as diabetic retinopathy and macular degeneration. Some areas of the retina with these conditions must be photocoagulated by laser to stop the progression of the disease and prevelant diseases. Deliminating these areas depends on the delineation of the lesions and anatomical structures of the retina. In this thesis we proposes a simple yet an efficient approach for automatic detection of the exudates of the Diabetic Retinopathy. The detection of exudates of diabetic retinopathy is composed of four main steps: 1. Max filtering of the fudus image converted to grayscale. 2. Fitting a polynomial curve composed of three line segments to the cumalative histogram and specified the second break level as a threhold level 3. Removing optic disk and false exudate regions from the image 4. Finally thresholding the image in the determined regions to get exudates. After exudate detection statistics of exudates have also been computed. The main contribution of this thesis is the automatic threshold level specification approach. The method is verified by an expert and it is seen that the proposed method is promising.
Trend removal of ecg signal with LMS algoritm
In this study, the LMS algorithm was applied to detect and track baseline introduced by involuntary movements during the acquisition of electrocardiogram signals, skin mismatch with electrode, or any movement caused by breathing or breathing of the patient (baseline wander). The baseline wander was removed from the signal and corrected. For this purpose, the baseline signal was described using two models: a varying constant and a line with a varying slope and constant. The latter is not found in the literature and is the contribution of this study. A fixed length window moved along the signal and the model parameters corresponding to the each segment were computed. The constant directly provided the baseline and the end of line of each window plotted the baseline wander. The results were reported by applying this approach to a simulated electrocardiogram signal and a real signal received from MIH-BIH database records. The outcome shows that modelling signal segment with a line adopts to the baseline faster than the constant model.
Categorization of normal and abnormal heart ryhtms from phonocardiogram signals
The phonocardiogram is based on recording the mechanical sounds of the heart, especially the sounds of the heart valves, and subsequent analysis. It provides an objective assessment of the cardiovascular system, which is vital for human life. In this thesis, the classification and analysis of normal and abnormal heart sounds obtained from the PCG signal was investigated. The data was obtained from challenge 2016 Physio-net database which consisted of two groups; A and B. Phonocardiogram signals was first decomposed into S1, S2 and S4 waves. Then bursts in each wave were segmented, and average power and dominant frequency were calculated from the segments of these signals. A simple linear Bayesian classifier was used to categorize heart sounds. The average accuracy of group A and B was obtained as 61% and 48% respectively. It was observed that the method fails for group B. However, the result for A was promising.
Olive tree crown detection, delineation and counting by using image processing techniques
UAVs are rapidly improving and increasing to enhance satellite based remote discovery. UAV (Unmanned Aerial Vehicle) remote sensing and low altitude remote sensing (LARS) applications play a significant role in the environment. This thesis shows the utilization of LARS in agriculture especially in the farming for detection and counting the "olive trees" by the proposed algorithms in this study. The image acquired by RGB camera and the processing of detection and counting was by utilizing Digital Image Processing techniques and Machine Learning algorithms. Several methods preformed to get better results, the first method utilized Uneven Illumination Correction, Gaussian filter, Standard Deviation filter and Circular Hough Transform (CHT) to detect olive trees. The second method utilized Histogram Equalization (HE), Mean, Median, and Wiener filters with 3 unsupervised machine learning algorithms (Clustering technique) are K-means, Fuzzy C–means (FCM), and Expectation Maximization (EM) algorithms and Morphological Operations then Circular Hough Transform. The third method implemented supervised machine learning algorithms (Classification) are K-Nearest Neighbor, Linear Discriminant Analysis, and Support Vector Machine.
İnsansiz hava aracı (İHA) görüntülerinin analizi ile karpuz meyvelerinin tespiti üzerine bir araştırma
Unmanned aerial vehicles (UAV) equipped with a digital camera are one of the technologies that the precision agriculture profits from. In this study, watermelons in the images obtained by a UAV from a watermelon field in Sarıçam, Adana, Turkey were segmented. For the study, three approaches were implemented. In the first approach, Haralick features and Bayes Linear Discriminant Analysis (LDA) methods with two categories were used. In the second approach, the first approach was combined with k-means clustering. Next, the second approach was developed by considering three categories with one versus all classifier in the final approach. The classification performance of each approach was evaluated and reviewed by utilising confusion matrices obtained from the classifier outcomes. The average categorization accuracy and the rate of detected watermelons without incorporating clustering outcome were 96.5% and 98.5% respectively. It is worth emphasizing that k-means clustering enhances the segmentation and, consequently, the accuracies. This study is the first step of yield estimation in watermelon production. It is believed that watermelon detection using image processing technology can be an asset to farmers and dealers regarding yield estimation and marketing. Key Words: UAV; image segmentation; Haralick features; linear classifier; k-means clustering; precision agriculture; watermelon detection.