Histopathological image analysis using deep learning

Cancer is an extremely dangerous disease that can be defined by the uncontrolled multiplication of one or a group of cells. According to recent studies, the deaths from cancer among the general deaths are in the first rank. Early diagnosis is crucial to prevent deaths from cancer. Early medical interventions increase the chance of survival. Biopsy is used as the gold standard for cancer diagnosis. In the biopsy procedure, samples taken from risky tissues are examined under appropriate microscopes by expert pathologists. Training and specialization of pathologists for the examination and decision-making of tissue fragments is a long and costly process. Even if they have successfully completed these processes, two different pathologists can make different decisions for the same piece of tissue. In order to eliminate this subjectivity, image processing methods that can calculate by using quantitative data are used. Early studies were based on obtaining the mathematically hand-determined properties from histopathological images. This feature extraction techniques, determined by image processing researchers, required a lot of experience and could not achieve high success for each image. In particular, the complex color information in the spatial structure of the tissue components in the histopathological images complicates this process. From the general framework, histopathological images include many color variations, pixel-level parasitic information, and many challenging factors, such as differences in cell lineage. Analyzing such image content with features that are not automatically determined is possible for a limited success level and a limited number of textures. In recent years, inspirational developments in artificial intelligence algorithms and rapid growth in computer hardware have been a glimmer of hope for automated feature extraction methods and have eliminated the negative effects arising from hand-crafted features. Thanks to its success in applications, deep learning method has become the most popular among automatic feature extraction and classification algorithms. Convolutional neural networks, which produce highly successful results especially for image processing problems, are used for images in almost all areas. Convolutional neural networks automatically remove features from the images and classify them by their own fully connected networks. According to the architectural structures formed, it can perform high level operations such as classification, segmentation and interpretation. In this thesis, a highly efficient and stable classification technique and a new semantic segmentation architecture are presented for the analysis of full-size histopathological images using convolutional neural networks. For this purpose, firstly the structure of histopathological images and the success of traditional methods on these images were examined. Then, the images were analyzed with linear convolutional neural networks and the effect of network parameters on the analysis success was investigated. In the light of all these data, a new convolutional neural network architecture has been proposed for the classification of histopathological images. The proposed architecture consists of a soft preprocessing layer for histopathological images and a convolutional neural network architecture of a linear structure. The proposed model is compared with other classifiers in the literature and appears to be more successful. As a result of the classification, the information generated by the proposed algorithm usually contains numerical information about large-scale areas of the image, which are cancerous and normal. However, the evaluation of an extremely important disease, such as cancer, by the artificial intelligence program is seen as insufficient. Such a decision is taken jointly by expert pathologists. It is generally the procedure to apply for the decision of more than one pathologist to make sure that one tissue is cancerous. Therefore, the results produced by artificial intelligence have an important deficiency to be a counseling system for pathologists. The results do not shorten the pathologist's review process because there is no field information and visuality. Semantic segmentation technique was used to overcome this problem. In the proposed system, the background, stained areas, cancerous cells and normal cells in histopathological images are markedly recognizable. During the development of this algorithm, the strongest semantic convolutional neural networks in the literature were examined and the strengths of these structures were determined. A new semantic segmentation architecture with all these strengths has been proposed. The proposed architecture produces visual results and allows the pathologist to easily understand the risk of cancer. In this way, the workload of pathologists is reduced and a quick consultation system is created. The proposed semantic segmentation architecture has been compared with other semantic segmentation algorithms and has been found to produce more successful results.