Development of extreme learning machine based classification algorithms for analysis of video images

Video analysis is of critical importance in medical science with regard to the microscopic-size intensive and sensitive data contained by the images. Video analysis can be defined as the process of obtaining meaningful and interpretable information about objects with moving object recognition and monitoring in videos recorded with microscopic cameras. In this thesis study, the use and development of Extreme Learning Machine (ELM) algorithms were performed in the analysis of video images. In order to test the performance of this developed hybrid (Hybrid-Differancial Evolation-Extreme Learning Machine) method, the correct classification of sperm motiles was performed on real sperm videos obtained. In medical literature, the primary cause of infertility problem in males who use sperms for fertilization is the low quality of sperms. The quality of sperm is determined with concentration (count), motility (movement) and morphology (structure) values. Specialists implement a spermiogram during diagnosis to detect the problem. This test consists of motility and morphology analysis and can be implemented using conventional or computer-assisted methods. In conventional motility analysis, specialists determine the sperm cell motility class by visual examination of seminal fluid under microscope. Computer-assisted methods, which is made up of software and hardware systems that detect and classify the sperm cell motility by video analysis. The aim of this thesis study is to detect feature parameters resulting from the motility of sperm cells with video analysis methods and to classify them with (ELM) classifier algorithms. To this end, the designed and applied system consists of three stages. In the first stage, Gauss Mixture Model which is one of the background/foreground segmentation methods was used for the recognition of each sperm cell in sperm motility videos and Kalman Filter-Hungarian Algorithm methods were used for monitoring them. In the second stage, sperm motility feature data were obtained by using coordinate information of each recognized motile sperm cell. In the final stage, motility feature data were classified with ELM and results were examined. Assessment of results shows that the performance of ELM method was higher compared to other classification methods (ANN, SVM and NB). What's new in this thesis study is that ELM's hidden layer cell counts and activation function selection process were optimized with the DE-ELM method developed by us, thereby increasing the performance of ELM. Furthermore, the applied software system was developed into a program package and aimed to use as artificial intelligence-based sperm motility. This aim contributes to the development of artificial intelligence-based systems on one hand and of diagnosis methods that will be clinically effective in the solution of infertility problem in men on the other.