Adnan Fatih Kocamaz

Continuous rotation invariant features for gradient-based texture classification

Four highly discriminative and continuous rotation invariant methods are proposed.We use the Hessian matrix and Gaussian derivative filters.Verified on the CUReT, KTH-TIPS, KTH-TIPS2-a, UIUC and Brodatz texture datasets. Extracting rotation invariant features is a valuable technique for the effective classification of rotation invariant texture. The Histograms of Oriented Gradients (HOG) algorithm has been proved to be theoretically simple, and has been applied in many areas. Also, the co-occurrence HOG (CoHOG) algorithm provides a unified description including both statistical and differential properties of a texture patch. However, HOG and CoHOG have some shortcomings: they discard some important texture information and are not invariant to rotation. In this paper, based on the original HOG and CoHOG algorithms, four novel feature extraction methods are proposed. The first method uses Gaussian derivative filters named GDF-HOG. The second and the third methods use eigenvalues of the Hessian matrix named Eig(Hess)-HOG and Eig(Hess)-CoHOG, respectively. The fourth method exploits the Gaussian and means curvatures to calculate curvatures of the image surface named GM-CoHOG. We have empirically shown that the proposed novel extended HOG and CoHOG methods provide useful information for rotation invariance. The classification results are compared with original HOG and CoHOG algorithms methods on the CUReT, KTH-TIPS, KTH-TIPS2-a and UIUC datasets show that proposed four methods achieve best classification result on all datasets. In addition, we make a comparison with several well-known descriptors. The experiments of rotation invariant analysis are carried out on the Brodatz dataset, and promising results are obtained from those experiments.

Cardiotocography analysis based on segmentation-based fractal texture decomposition and extreme learning machine

Fetal heart rate (FHR) has notable patterns for the assessment of fetal physiology and typical stress conditions. FHR signals are obtained using cardiotocography (CTG) devices also providing uterine activities simultaneously and fetal movements. In this study, a total of 88 records consisting of 44 normal and 44 hypoxic fetuses instances obtained from publicly available CTU-UHB database have been considered. The basic morphological features supporting clinical diagnosis, the powers of 4 different spectral bands and Lempel Ziv complexity have been used to define FHR signals. Also, it has been proposed to use segmentation-based fractal texture analysis (SFTA) to identify the signals more accurately. The obtained feature set was applied as the input to extreme learning machine (ELM) with 5-fold cross-validation method. According to experimental results, 79.65% of accuracy, 79.92% of specificity, and 80.95% of sensitivity were obtained. It was observed that the SFTA offers useful statistical features to distinguish normal and hypoxic fetuses.

Cardiotocography signals with artificial neural network and extreme learning machine

Cardiotocography (CTG) is a monitoring technique that is used routinely during pregnancy and labor to assess fetal well-being. CTG consists of two signals which are fetal heart rate (FHR) and uterine contraction (UC). Twenty-one features representing the characteristic of FHR have been used in this work. The features are obtained from a large dataset consisting of 2126 records in UCI Machine Learning Repository. The prominent features, such as baseline, the number of acceleration and deceleration patterns, and variability recommended by International Federation of Gynecology and Obstetrics (FIGO) have also taken into account during CTG analysis. The features were applied as the input to feedforward neural network (ANN) and Extreme Learning Machine (ELM) to classify FHR patterns in this study. FHR is recently divided into three classes as normal, suspicious and pathological. According to the results of this study, the accuracy of classification of ANN and ELM were obtained as 91.84% and 93.42%, respectively.

Using wavelet transform for cardiotocography signals classification

As a fetal surveillance technique, cardiotocography (CTG) involves fetal heart rate (FHR), uterine contraction activities, and fetal movements. CTG is practiced as a primary diagnostic test throughout the world to identify events that may pose a risk to the fetus during pregnancy and delivery. In this work, FHR signals carrying vital information on fetus were analyzed by using Haar (haar), Daubechies (db5), and Symlets (sym5) mother wavelet families between levels 1 and 12. The traditionally used morphological and linear features are obtained from FHR. Also, p-norm, Frobenius form, infinity, and negative infinity norms which are obtained separately from the each of the wavelet components were used as a feature to support the classification. The obtained features were applied as an input to k-nearest neighbors (kNN) and artificial neural network (ANN) classifiers in order to discriminate the normal and hypoxic fetuses. According to experimental results, 90.51% and 90.21% classification success on the discrimination of normal and hypoxic fetuses were achieved by using haar at level 4 and kNN.

A novel software for comprehensive analysis of cardiotocography signals “CTG-OAS”

The research interest in fetal heart rate (FHR) monitoring dates back to the 1960s, and the breakthrough on fetal surveillance has been seen during the 1990s with computerized systems. Notwithstanding the general use of cardiotocography (CTG) in fetal monitoring, the assessment of fetal well-being exhibits a significant inter- and even intraobserver variability. Computerized CTG analysis has seen as the most promising way to tackle of the main shortcomings of visual CTG assessment. In this study, a novel software developed for research purposes is introduced. The software named as CTG Open Access Software (CTG-OAS) characterizes FHR signals by using comprehensive features obtained from different fields such as morphological, linear, nonlinear, time-frequency, discrete wavelet transform, and image-based time-frequency domains. The software also covers the main procedures which are necessary for the context of machine learning. More specifically, CTG-OAS presents several tools for performing the preprocessing, feature extraction, feature selection, and classification processes. The proposed software was practiced on CTU-UHB database with 552 raw CTG samples. In addition, a case study with Support Vector Machine classifier was performed in the study via CTG-OAS. According to experimental results, statistical parameters were obtained as accuracy equal to 87.97%, sensitivity equal to 89.04%, specificity equal to 81.36% and, quality index equal to 85.11%.

Robot control with graph based edge measure in real time image frames

Robotic path plan extraction is one of the major study area focused with image processing techniques except from estimation based methods, in real time robotic systems. In this study; I. First phase: It is aimed to track a mobile robot and target point by detecting start and target points in real time image frames, continually. II. Second phase: It is aimed to exhibit a novel sensor free kinematic model to determine power ratio transferred to the left and right wheels of the robot with decision tree method by utilizing a graph based method on detected object points. Right and left wheel position, robot center and labeled target positions are acquired with thresholding method by using labels placed on robot. A graph has been created on an image by admitting all detected areas as nodes. The processes of orientating and delivering mobile robot to the target position has been modelled according to distance values of edges between wheels and target.

Vision-based decision tree controller design method sensorless application by using angle knowledge

In this study, a new control model for differential drive mobile robots was developed by using image-based decision tree method(DTM). Developed mobile robot control model was designed in an obstacle-free environment. The wheel encoder sensor was designed as a controller capable of independent positioning by using real-time images from overhead cameras on the birds eye view. In this new method, a virtual triangular area between the target and the robot was created. These triangular base angles were calculated on the image. Decision tree controller was determined as the difference between the base angles by branching. Decision Tree leaves control determines the left and right wheel speeds depending on the difference model design was carried out. The developed new controller model was tested on Khepera IV robot. In practice, the robots speed and angle of the body was carried uncensored control and it was observed to find the target in different applications. Application of the results and performance of the system was shown.

Prognostic model based on image-based time-frequency features and genetic algorithm for fetal hypoxia assessment

Cardiotocography (CTG) is applied routinely for fetal monitoring during the perinatal period to decrease the rates of neonatal mortality and morbidity as well as unnecessary interventions. The analysis of CTG traces has become an indispensable part of present clinical practices; however, it also has serious drawbacks, such as poor specificity and variability in its interpretation. The automated CTG analysis is seen as the most promising way to overcome these disadvantages. In this study, a novel prognostic model is proposed for predicting fetal hypoxia from CTG traces based on an innovative approach called image-based time-frequency (IBTF) analysis comprised of a combination of short time Fourier transform (STFT) and gray level co-occurrence matrix (GLCM). More specifically, from a graphical representation of the fetal heart rate (FHR) signal, the spectrogram is obtained by using STFT. The spectrogram images are converted into 8-bit grayscale images, and IBTF features such as contrast, correlation, energy, and homogeneity are utilized for identifying FHR signals. At the final stage of the analysis, different subsets of the feature space are applied as the input to the least square support vector machine (LS-SVM) classifier to determine the most informative subset. For this particular purpose, the genetic algorithm is employed. The prognostic model was performed on the open-access intrapartum CTU-UHB CTG database. The sensitivity and specificity obtained using only conventional features were 57.33% and 67.24%, respectively, whereas the most effective results were achieved using a combination of conventional and IBTF features, with a sensitivity of 63.45% and a specificity of 65.88%. Conclusively, this study provides a new promising approach for feature extraction of FHR signals. In addition, the experimental outcomes showed that IBTF features provided an increase in the classification accuracy.

Fetal Hypoxia Detection Based on Deep Convolutional Neural Network with Transfer Learning Approach

Electronic fetal monitoring (EFM) device which is used to record Fetal Heart Rate (FHR) and Uterine Contraction (UC) signals simultaneously is one of the significant tools in terms of the present obstetric clinical applications. In clinical practice, EFM traces are routinely evaluated with visual inspection by observers. For this reason, such a subjective interpretation has been caused various conflicts among observers to arise. Although the existing of international guidelines for ensuring more consistent assessment, the automated FHR analysis has been adopted as the most promising solution. In this study, an innovative approach based on deep convolutional neural network (DCNN) is proposed to classify FHR signals as normal and abnormal. The proposed method composes of three stages. FHR signals are passed through a set of preprocessing procedures in order to ensure more meaningful input images, firstly. Then, a visual representation of time-frequency information, spectrograms are obtained with the help of the Short Time Fourier Transform (STFT). Finally, DCNN method is utilized to classify FHR signals. To this end, the colored spectrograms images are used to train the network. In order to evaluate the proposed model, we conducted extensive experiments on the open CTU-UHB database considering the area under the receiver operating characteristic curve and other several performance metrics derived from the confusion matrix. Consequently, we achieved encouraging results.

Open-access software for analysis of fetal heart rate signals

Abstract Cardiotocography (CTG) comprises fetal heart rate (FHR) and uterine contraction (UC) signals that are simultaneously recorded. In clinical practice, a visual examination is subjectively performed by observers depending on the guidelines to evaluate CTG traces. Owing to this visual assessment, the variability in the interpretation of CTG between inter- and even intra-observers is considerably high. In addition, traditional clinical practice involves different human factors that distort the quantitative quality of the evaluation. Automated CTG analysis is the most promising way to tackle the main shortcomings of CTG by providing reproducibility of the evaluation as well as the quantitative results. In this study, open-access software (CTG-OAS) developed with MATLAB® is introduced for the analysis of FHR signals. The software contains important processes of the automated CTG analysis, from accessing the database to conducting model evaluations. In addition to traditionally used morphological, linear, nonlinear, and time–frequency features, the developed software introduces an innovative approach called image-based time–frequency features to characterize FHR signals. All functions of the software are well documented, and it is distributed freely for research purposes. In addition, an experimental study on the publicly accessible CTU-UHB database was performed using CTG-OAS to test the reliability of the software. The experimental study obtained results that included an accuracy of 77.81%, sensitivity of 76.83%, specificity of 78.27%, and geometric mean of 77.29%. These fairly promising results indicate that the software can be a valuable tool for the analysis of CTG signals. In addition, the results obtained using CTG-OAS can be easily compared to different algorithms. Moreover, different experimental setups can be designed using the tools provided by the software. Thus, the software can contribute to the development of new algorithms.