Mitosis detection in histopathological images

In the last decades extremely important developments have been realized with the advance of technology in digital histopathology. Now, it is possible to make Computer-Aided Diagnosis (CAD) on histopathological images with advances in high-resolution imaging technology. Various algorithms from signal and image processing, pattern recognition and machine learning methods have been applied to histopathological images. Several studies have been done especially on mitosis detection stage in histopathological images, which plays major role in diagnosis of cancer. With the research and development (R&D) studies on Computer-Aided Diagnosis (CAD) in universities and the private sector have reached promising results on cancer diagnosis and prognosis. The main purposes of these studies are to reduce the workload of pathologists and the impact of errors caused by human factor. The first stage in digital histopathology is to separate cellular structures from non-cellular structures. Those cellular structures are extracted by applying pattern recognition and machine learning methods to the images taken from high resolutional scanners. The cell sizes, boundaries, distributions and shapes are very important for a good detection of cellular structures. After the separation of cellular structures, mitotic and non-mitotic cells must be determined as much as possible. Feature extraction stage is a very important step to distinguish the mitotic cells from non-mitotic ones. In this context; in the first approach of this thesis study, the feature extraction stage has been realized by using Histogram of Oriented Gradients (HOG) which is commonly used for human detection. In the second approach, Haralick texture descriptor method is used with making use of spatial relevance. In this method, feature vectors of each specific pixel are extracted by gray level co-occurrence matrices (GLCM) of a predefined pixel region. After feature extraction stage, the classification process is performed with supervised learning algorithms; Support Vector Machines (SVM), Random Forest (RaF), and Rotation Forest (RotF) algorithms. Histopathological images have been also transformed to La*b* color space besides RGB color space. The classification result has presented comparatively in tables. In the second approach, the classification is performed with random forest (RaF) algorithm. Both the effects of RGB and La*b* color spaces to the classification result have compared pairwise. But it is proved that color spaces have little effects on classification results in the favor of La*b* colorspace.