Theses supervised by Prof. Dr. Fatih Vehbi Çelebi
19 theses · Ankara Yıldırım Beyazıt University
Early detection of lung cancer
The clinical diagnostics for lung cancer are mostly depended on physical and biochemical techniques. Imaging processes in computed tomography (CT) screening (low-dose computed tomography - LDCT) is convenient for discovering lung cancer in the early stages. CT scan images for this study were obtained from Ankara Atatürk Training and Research Hospital and also from the online international database of "The Lung Image Database Consortium (LIDC)". Correct decision for diagnosis of lung cancer using CT scanning requires some processes to remove noise from image in enhancement stage. The noise removing process in this thesis have been proposed using a gradient magnitude in sobel filter. Finding edges on the images is a first step to detect nodule in the tissues. As well as following morphology operations to isolate background from the foreground is very important because background image represents the tissue of lung to locate a tumor on it, therefore foreground is unnecessary. Second part in the thesis includes usage of watershed algorithm for segmentation of tumor from the tissue. Labeling nodular irregular area inside the tissue leads to over-segmentation on image by connected components and markers which have different values as intensity values and regional minimum represented in the foreground. Markers classify tumor area by labeling high intensity values, locating the region of interest in normal image for cutting random area from the tissue. Distinguishing the normal and abnormal images that needs to use statistical methods is depended on the cancer type. To get feature extraction of nodule shape provides certain parameters which is an essential step for classification processes. Last part in this thesis focuses on classification processes on a dataset which includes values belong to five parameters that were taken from statistical method results. This dataset involves 306 images consisting of 153 normal images and 153 abnormal images. This database is implemented in nine classification algorithms and different results of accuracy performance were taken according to the theory of these algorithms. This study aims detecting tumor area and getting high performance accuracy after classification were done to find out the best algorithm to help the doctor for the diagnosis. Keywords: Lung cancer, early detection, sobel filter, morphology operations, watershed algorithm
The clustering of public hospitals for the productivity scorecard application
Identifying and Grouping the Roles of Hospitals on an Institutional Basis is one of the study of Ministry of Health. One of the consequences of the restructuring was the publication of Decree Law No. 663. In this context hospitals which are affiliated to Turkey Public Hospital Institution required to be evaluated in 6-month or annual period with this Legislative Decree. As required to this article for evaluation "The Productivity Scorecard Application" has started based on "The Balanced Institutive Scorecard Model". For each indicator, which is involved in "The Productivity Scorecard Application", acceptable values have been determined using different methods. In some indicator cards, acceptable value has been regarded as the average of the similar hospitals service classes. In the study presented in this thesis; it is focused on the study of k-Means clustering algorithm of data mining techniques and hospitals which are affiliated to Turkey Public Hospital Institution and clustering of hospitals whose productivity scorecard will be estimated. Hospitals have been clustered in terms of their financial status, equipment capacity, staff capacity and produced medical services' volume and variety. As a result of clustering work carried out within the scope of this study; 597 hospitals to be assessed by The Productivity Scorecard Application were divided into 16 clusters. 149 attributes have been determined for the hospital data used as input in the clustering algorithm and data has been collected from the data of year 2016. The validity of the clustering results was tested by taking expert opinions and also evaluated according to the distribution ratios of the hospital roles formed in the clusters. After these evaluations, it can be clearly stated that the clustering study has been found to be successful to substantially, taking into account the systematic gathering of data from the application resources and the detection of right outliers. Keywords: Hospital clustering, k-means clustering algorithm, bisecting k-means, productivity scorecard application.
Kuantum kaskat lazerlerin karakteristik parametrelerinin üç boyutlu modellenmesi
Quantum cascade lasers (QCL) are semiconductor lasers that contain quantum wells. Quantum wells provide distinct quality for QCLs. Changing width of quantum wells cause to obtain different wavelengths for QCLs. QCLs have critical characteristic quantities such as optical gain, linewidth enhancement factor (LEF) and refractive index change. These parameters differ according to conditions. In this study; relating to injection current and wavelength the changes of optical gain, LEF (Alpha parameter) and refractive index change which are characteristic parameters of QCLs are modelled three dimensionally considering experimental values. Surface fitting techniques: regression analysis (lowest, polynomial) and method of least squares are applied to provide the optimal surface. Both the training and test results are used to obtain surface curves of characteristic quantities with minimum error. Except method of least squares, MATLAB program is used to find the surfaces and the errors of techniques. In this study, we present the best fitting technique to find the ideal parameters for each different QCLs' variables.
A smart energy-efficient fire prediction system for a charcoal manufacturing plant
In recent years, there has been a tendency to make objects, buildings, and cities connected and smart. The themes of smart buildings, cities, and grids are new and behind these concepts, there is a digital revolution that is transforming all the sectors of society, including agriculture, finance, logistics, etc. It is the case of factories that use all the technologies that can allow them to operate efficiently while producing good performance. However, regardless of the technologies used, work in the factories presents risks, in particular fire risks for many reasons. For example, a charcoal manufacturing plant presents fire risks. In this regard, many technologies, with the help of the appropriate equipment are nowadays used to predict fires. For these equipments to have a long life, it is necessary to properly manage the energy with which they operate. For this reason, to ensure better energy management, a smart energy-efficient fire prediction system for a charcoal manufacturing plant is proposed. Problems related to excessive energy consumption by active equipment in a network need to be resolved. A dataset has been used to train and test data to predict the risk of fire with the help of four different machine learning algorithms. We also tried to optimize the energy consumption of the network by using Genetic Algorithm (GA) and SARSA algorithm, a type of reinforcement learning algorithm.
Redesigning fixed-wing loitering munition as solar supported by using reverse engineering method
Conventional warfare has been transformed into hybrid warfare or asymmetric threat in recent decades. Due to this reason, the world's armed forces are constantly taking innovative steps to improve their defense technologies. Thanks to this technological development, armed forces all over the world are begun to equip with a variety of UAVs in accordance with their military mission. One of these UAVs, supplanted in army inventories, loitering munitions, also known as kamikaze drones, which create an explosive effect on the target by hitting the in defilade position, are autonomous weapon systems that operate on the "search-find-destroy" principle. In this thesis, first the history of solar-powered manned/unmanned aerial vehicles (UAVs) all over the world, , the development process of UAV technology in Turkey and classification of UAVs and the place of loitering munitions will be explained in details. Then, one of the fixed-wing loitering munitions, ALPAGU which are effectively used in today's battlefields will be taken on board and its working principles, features and specifications will be explained. Subsequently, taking into account the irradiance values of the Şırnak province, it will be calculated whether the flight duration of 15 minutes can be extended by reverse engineering method with the additional solar support components. Whether its cost-efficient or not, this study is the first of its kind to apply a solar support system to a loitering munition in order to extend its flight endurance by harnessing solar energy. Finally, as a result of the calculations made on regarding the solar support components applied to the ALPAGU loitering munition, it will be theoretically proven that the time in the air can be extended by 11.2%, in numerical terms, 1.76 minutes (105.6 seconds). Keywords: Loitering Munitions, Kamikaze Drones, Suicide Drones, Solar Cells, Solar Energy, Solar Supported, UAVs, MPPT, LiHV Batteries
Parkinson's disease diagnosis by using autoencoder based on deep neural network (DNN) and metaheuristic method
Parkinson's disease (PD) is a neurodegenerative disorder and affects the nerve cells that produce dopamine in the brain. In this thesis, we investigated comparative studies on the different scenarios such as AutoEncoder and Ant Colony Optimization feature selection algorithms to determine the effective features in the diagnosis of Parkinson's disease. These algorithms are implemented to the voice obtained from an online repository. Then selected features are used with the Decision tree, SVM, K-NN, Ensemble, Naive Bayes and Discriminant classifiers for each of the binary classification problems. The proposed methods are evaluated with the sensitivity, specificity, precision, recall and accuracy criteria. The suggested systems are trained and tested with these classifiers separately to carry out a comparative study and to analyze the success of feature selection methods in discriminating healthy people and PD patients. In Parkinson's Disease Dataset 24 features were obtained from the signal voices. Some of the features in the training of the classifier have problems and these problems reduce the accuracy of the system. It is found that for K-NN and Ensemble classification methods both Ant Colony Optimization (ACO) and Autoencoder have the same and the best training performance. Testing results show that the accuracy rate of the improved ACO is higher than the Autoencoder method. Keywords: Improved ant colony optimization, autoencoder, deep learning, feature selection, parkinson's disease.
Tool-path optimization in FFF 3D printing machine using lines approach algorithms
By using Fused Filament Fabrication (FFF) 3D Printing Machines we can create simple to complicate objects by printing successive thin layers. The printing tool formed each layer by moving on a selected path. According to the fact that there are a huge number of layers to print, the selection of path determines the overall time needed for producing an object. Therefore, reducing the length of paths becomes an essential part of the printing process. In this study, we presented different algorithms to optimize each layer, they are LinearGreedy and LinearTwoOpt, by covert each layer into lines instead of points to find the optimal path. Hybrid algorithms are also presented by selecting an initial solution with a fast algorithm (LinearGreedy) and after that to try to improve it by further exploration of the solution space with the other one (LinearTwoOpt).
Development of sorting and searching algorithms
The increase in the rate of data is much higher than the growth in the speed of computers, which results in a heavy emphasis on sort and search algorithms in the research literature. In this thesis, we propose a new efficient sorting algorithm based on the insertion sort concept. The proposed algorithm is called Bidirectional Conditional Insertion Sort (BCIS). It is an in-place sorting algorithm, and it has a remarkably efficient average case time complexity when compared with a standard insertion sort. By comparing our new algorithm with the QuickSort algorithm, BCIS shows faster average case times for relatively small arrays of up to 1500 elements. Furthermore, BCIS was observed to be faster than QuickSort within the high rate of duplicated elements even for large arrays. We present a hybrid algorithm to search ordered datasets based on the idea of interpolation and binary search. The presented algorithm is called Hybrid Search (HS), which is designed to work efficiently on unknown distributed ordered datasets. Experimental results showed that our proposed algorithm has better performance when compared with other algorithms that use a similar approach. Additionally, this study describes and analyses an unaddressed issue in the implementation of the binary search. In spite of this issue not affecting the correctness of the algorithm, it decreases its performance. However, the study presents a precise analytical approach to describe the behavior of the binary search in terms of comparisons number. With the help of this method, the complexity of the weak implementation is proved. Experimental results show that the weak implementation is slower than the correct implementation when a large sized search key is used. The presence of this implementation issue within other algorithms is also investigated. Finally, we present two efficient search algorithms, the first of which is an improved implementation of the ternary search, and the second a new algorithm called Binary-Quaternary search (BQ search). BQ search uses a new efficient divide-and-conquer technique. Both proposed algorithms, theoretically and experimentally, show better performance when compared with binary searches. Although the proposed BQ search displays a slightly higher average comparisons number than the improved ternary search, experimentally the BQ search shows better performance compared with improved ternary searches under some conditions.
Steerable filter in frequency domain with rose curve
Feature extraction in image processing area is a vital part for further processes in this area. Feature extraction from images is mainly conducted by the filters both in spatial domain and frequency domain. Finding suitable, especially when directional features are in question, in spatial domain is a difficult task. However, feature extraction in frequency domain is more flexible since directions are more obvious in this area. For feature extraction in frequency domain, several methods have been proposed and actively been in use image processing area. One of these methods is Contourlet Transform which does not offer easy direction selectivity. In other words, directions are constant in Contourlet Transform. Another widely used method is 2D Gabor Filter. Although 2D Gabor Filter has flexibility in direction selection, the method suffers from large number of parameters to decide. In this thesis study, a novel filter for feature extraction in frequency domain using Rose Curve has been proposed. One of the most important properties of the proposed method is angle selectivity. Easiness for the direction selectivity has been achieved by this method. Another property of the proposed method is the number of parameters to decide. The proposed method has only three, easily adjustable parameters. The proposed method has been evaluated on the facial expression classification problem and higher results in terms of different metrics than some of the state of the arts methods in the literature were obtained.
A new buffer management solution to improve the performance of mac layer in wireless sensor networks
Wireless Sensor Network (WSN) is an autonomous network that detects physical changes in an environment and reports the relevant information to a central point for further investigation. WSNs are an important type of resource-constrained distributed systems. There are many limitations on the sensor nodes such as: energy, memory, computation capability, storage, transmission range and etc. All these limitations make efficiency and effectiveness very important and highly demanded features for WSNs. In this study, buffer management mechanisms that can be utilized on sensor nodes at different communication layers were examined using mathematical analyses and simulations. In this context, a new mathematical model was introduced and various number of simulations were performed in order to show whether conventional buffer management solutions are appropriate for WSNs. The results obtained from simulations were compared with each other and the proposed our new mathematical model. Through the investigations of buffer management solutions for WSNs the usage of memory and its optimality was tried to be come out. It was found that conventional buffer management solutions are not appropriate for WSNs and they don't satisfy WSN specific needs. The utilization of buffer either would be sub-optimal or the prioritization between packets could not be supported in these buffer management approaches. There is a clear need a new approach for buffer management in WSNs. In here, we propose a novel buffer management solution that improves the general performance of Medium Access Control (MAC) layer plans, in particular those crafted for WSNs. The success of our proposed solution were discussed and it was compared with other well-known buffer management solutions. The comparison results in different plots are presented in this study.
A novel hybrid approach to chan-vese algorithm for deformable contour based image segmentation
Image segmentation process is the most important and difficult step in object recognition systems. A number of object segmentation methods based on deformable models, also known as active contour models, and which are used to find object boundaries on images have been proposed. However, the main definitions of these methods depend on: the contour initialization and the correct convergence of the subsequent solution. When the initial contour is not properly positioned, this causes to produce unsuccessful results due to problems such as the inability to converge to the desired result within the expected number of iterations, or getting stuck in local minima during the minimization of the energy function. In this thesis study, an image segmentation method based on the gravitational search algorithm which is a heuristic method, and on the active contour without edges model (Chan-Vese) has been developed to overcome the problem of the contour initialization. The proposed model has been developed by combining the gravitational search algorithm and the Chan-Vese algorithm in a hybrid way. It has been tested on various images, including some images from the Weizmann database and medical images. With the robust structure against the initial contour selection of the proposed model, more efficient results were obtained than the conventional Chan-Vese algorithm. Performance of the novel model has been evaluated in terms of accuracy and efficiency compared with the conventional model.
Design of a low phase noise lc voltage controlled oscillator based on native transistors
One of the most preferred LC VCO topologies is complementary LC VCO. The topology has difficulty in providing sufficient current to the oscillator core in the current controlled oscillators due to the low voltage headroom. A fully integrated 7.5 GHz VCO is designed and simulated using TSMC 65 nm CMOS technology using native transistors for the biasing circuit to overcome this problem. Measured phase noise is -121 dBc/Hz for 1 Mhz offset at the center frequency. The oscillator has %5.58 tuning range. It draws 1.8 mA current from a 1.2 V supply. 2 dB better phase noise performance is obtained compared to peer circuit used normal transistors in biasing circuit.
A graph database application to analyse social networks
The aim of the thesis is finding suspicious persons or vessel behaviors via Social Network Analysis using Neo4j as Graph database. To detect relationship between suspicious individuals and vessels by building a visual relation network with graph mining study to analyses social networks in marine vessels data for intelligence departments. For example, in the x-ship, 50 people were employed, were they involved in crime, or did they work on ships associated with another crime? To deal with this question drawing a visual network show the answers. It is aimed to create a database collected through the relational database and that includes crime events and related person data on the sea in the form of social networks and view the social networks on this database. SNA provides a roof for the simulation and presentation of a case which interacting units and their relationships. It involves a group of methods and tools to data collection, classification, pattern identification, prediction and visualization. There are limitations of crime networks like accessing the data is restricted reason of the sensitivity of the information and the "innocence factor", meaning that a person who has been suspected or charged for a crime might in fact be innocent and been registered wrongly. Due to these restrictions, the comments of individuals on involvement in crime have been transferred to law-enforcement. In parallel with this purpose, by choosing graph database model that gains a major advantage over traditional databases in social network modelling, an event-person database was built in the Neo4j platform and it is aimed to circulate the relevant authorities on the network by designing a web interface.
Introduction and benchmark result comparison of socially inspired algorithms
Socio-inspired algorithms are a particular type of metaheuristic that bases its behavior on human society. It is a simple approach to the complex world of evolutionary computing, with multiple simulations with recognizable patterns day by day. The aim of this study is to check whether these algorithms have the functional capacity to solve optimization problems. Through meticulous experimental analysis, the results of these algorithms will be collected for a known benchmark and compared with those of some reference algorithms that allow assessing their potential in this type of problem adequately.
A study on performance evaluation of optimization algorithms in the shortest path problem
Finding the shortest or least costly path between two points in the shortest time is crucial in many areas. The shortest path problem aims to move from a starting node on a graph to the destination node via the shortest path. Nowadays it is widely used in traffic applications, routing of internet traffic and programming of games. As time went by and with the development of science and technology, many scientists made researches to solve the shortest path problem, and many algorithms were developed. These algorithms address different kinds of problems and the shortest paths between nodes can be found with them. It can be observed that in the literature there are many and different studies on the shortest path problem. The aim of this study is to analyze and compare four of the algorithms used in shortest path problem. Dijkstra, Bellman – Ford, Johnson's and Floyd-Warshall Algorithms are displayed on the graph drawing application. It is discussed that which algorithm is better and more effective in finding the shortest path. The comparison of four different algorithms for positive weighted, non-directional and fully connected graphs on the same graph drawing application is remarkable because there is no study on this subject with the same spesifications. According to the results of the study, it is understood that algorithm that will be used should be selected in line with type of graph and problem. It is proved that Dijkstra algorithm is absolutely must be used for fully connected, positive weighted and non-directional graphs. Keywords: Graph, the shortest path, optimization, Dijkstra, Floyd-Warshall, Bellman-Ford, Johnson's
Investigating defacer behavior and defacement attacks using twitter
Many web-based attacks have been studied to understand how web hackers behave, but web site defacement attacks (malicious content manipulations of victim web sites) and defacers' behaviors have received less attention from researchers. This research fills this research gap via computational data-driven analysis of a public database of defacers and defacement attacks and activities of 96 selected defacers who were active on Twitter. We conducted a comprehensive analysis of the data: an analysis of a friendship graph with 10,360 nodes, an analysis on how sentiments of defacers related to attack patterns, and a topical modeling based analysis to study what defacers discussed publicly on Twitter. Our analysis revealed a number of key findings: a modular and hierarchical clustering method can help discover interesting subcommunities of defacers; sentiment analysis can help categorize behaviors of defacers in terms of attack patterns; and topic modeling revealed some focus topics (politics, country-specific topics, and technical discussions) among defacers on Twitter and also geographic links of defacers sharing similar topics. We believe that these findings are useful for a better understanding of defacers' behaviors, which could help the design and development of better solutions for detecting defacers and even preventing impeding defacement attacks.
Quantum key distribution protocol-based image encryption algorithm
In this thesis, a simulation of BB84 QKD protocol will be implemented to generate a secret key which is used for the encryption and decryption of digital images. Moreover, a method for combining the original image with the generated secret key will be proposed. In the proposed method, both the original image and the key are split into their RGB components, and a combination of XOR operations and modular arithmetic is applied to combine them. The QKD-based encryption method is tested on various standard test images of different dimensions. To ensure the robustness, efficiency, and performance of this QKD-based encryption scheme, several tests and analysis will be run on the resulting data including standard attack checks, correlation coefficient analysis, entropy analysis, histogram analysis. Finally, the results will be compared with the E91 QKD protocol. As the future of the secure-deemed public-key cryptography is in jeopardy, Quantum Key Distribution (QKD), which is part of Quantum Cryptography (QD), addresses these security concerns by providing a robust and secure solution. QKD utilizes the principles of quantum mechanics and the laws of physics to establish a secure secret key between two parties with the help of Heisenberg uncertainty principle and the no-cloning theorem. The key security feature of QKD protocols is the ability to detect whether an eavesdropper is present since an attempt to eavesdrop would cause a disturbance in the system and alert the original parties. There has been remarkable development regarding the design of QKD protocols recently, and many in-field prototypes are being implemented.
Hemositometre üzerindeki lösemi kanser hücrelerinin otomatik segmentasyonu
Cell counting is used to determine cell number and cell density as a key step in the laboratory workflow. Determining cell density is very important both for accurate and precise diagnosis of diseases and drug experiments. Due to the high cost of the cell counting machines, specialists often count cells manually with a hemocytometer. Counting cells manually is both a laborious and time-consuming activity. In this study, semantic cell segmentation method based on deep learning is presented to count cells automatically. The data set that is analyzed in this study contains 468 light microscope images of HL60 leukemia cancer cells on the hemocytometer that are taken from cell culture. 421 of 468 images in the image set were reserved for use with k=5 fold cross-validation and 47 for model validation. Pixel accuracy and mean intersection over union (IoU) metrics were used to evaluate the training performance of the model built with U-Net. Hyper-parameters optimization was applied by Grid Search Algorithm. As a result, average pixel accuracy was achieved 98 percent and average IoU 87 percent. There are 636 cells in the test images and the number of cells acquired by using connected component analysis from the segmented results is 511. Thus, the cell detection rate was achieved 80 percent. By transferring the method developed in this study into application, experts can carry out cell counting procedure automatically without using a costly cell counting machine. Therefore, it is thought that the presented method will contribute both time and budget saving.
Early detection of lung cancer
The clinical diagnostics for lung cancer are mostly depended on physical and biochemical techniques. Imaging processes in computed tomography (CT) screening (low-dose computed tomography - LDCT) is convenient for discovering lung cancer in the early stages. CT scan images for this study were obtained from Ankara Atatürk Training and Research Hospital and also from the online international database of "The Lung Image Database Consortium (LIDC)". Correct decision for diagnosis of lung cancer using CT scanning requires some processes to remove noise from image in enhancement stage. The noise removing process in this thesis have been proposed using a gradient magnitude in sobel filter. Finding edges on the images is a first step to detect nodule in the tissues. As well as following morphology operations to isolate background from the foreground is very important because background image represents the tissue of lung to locate a tumor on it, therefore foreground is unnecessary. Second part in the thesis includes usage of watershed algorithm for segmentation of tumor from the tissue. Labeling nodular irregular area inside the tissue leads to over-segmentation on image by connected components and markers which have different values as intensity values and regional minimum represented in the foreground. Markers classify tumor area by labeling high intensity values, locating the region of interest in normal image for cutting random area from the tissue. Distinguishing the normal and abnormal images that needs to use statistical methods is depended on the cancer type. To get feature extraction of nodule shape provides certain parameters which is an essential step for classification processes. Last part in this thesis focuses on classification processes on a dataset which includes values belong to five parameters that were taken from statistical method results. This dataset involves 306 images consisting of 153 normal images and 153 abnormal images. This database is implemented in nine classification algorithms and different results of accuracy performance were taken according to the theory of these algorithms. This study aims detecting tumor area and getting high performance accuracy after classification were done to find out the best algorithm to help the doctor for the diagnosis. Keywords: Lung cancer, early detection, sobel filter, morphology operations, watershed algorithm