Classification of ovarian follicles with deep learning

Microscopic images are used in the diagnosis and scientific studies of many diseases. Histological analyses of these images are performed by experts. In general, histological analysis refers to studies on cells, cell structures, and tissue. The increase in the number of images in the histological analysis takes a long time for the expert and also creates an excessive workload. Furthermore, the analyzes are often subjectively based on the knowledge of each specialist. Considering such problems, there is a need for intelligent systems that can perform objectively image analysis in less time. Different algorithms have been developed for the analysis of medical images with machine learning methods in the literature. However, these methods require the expert to process raw data. Deep Learning (DL) models have been developed for automatic feature discovery on raw data without the need for experts. DL is a sub-branch of artificial intelligence and is referred to as a generic name for deep network architectures. The error rate in object classification has declined sharply using DL models. The main reason for this success in DL can be explained as the discovery of different features of data in different deep layer structures. For this reason, in recent years, DL methods have been used frequently in many different fields, especially the analysis of medical images. In this thesis, a new method based on DL for segmentation and classification of ovarium images has been proposed. The proposed method was used for automatic counting of five different follicles that primordial, primary, preantral, secondary and tertiary, belonging to ovary tissue. In the literature, the automatic counting of these follicles was performed for the first time within the scope of this thesis. The proposed method consists of two parts: training and testing. In the training section, different Convolutional Neural Networks (CNN) were designed. In the test process, both segmentation and classification process is performed using trained CNN models and it is decided where and what a cell or cellular structure in the image. In the first part of the method, 55 different CNN models were designed. 43 of these models were designed for segmentation and 12 for classification. An original data set consisting of 10500 images of ovarian tissue was created to be used in the training and testing of these models. In this dataset, follicles in each image are labeled. Different data sets have also been created to train CNN models designed for segmentation and classification from labeled images. The success of the CNN model, which achieved the highest accuracy among the models for segmentation was 87.1%. The success rate of the highest accuracy CNN model for the classification process was 96.01%. In the test section of the method, filter-based segmentation is performed using CNN, which is trained primarily for segmentation. A new method has been proposed to remove the noise generated after this process and to determine the boundaries of cellular structures. Pretrained CNN model is used to classify the cellular structures whose boundaries are defined. In order to increase the segmentation accuracy of the developed method, General Stride (GS), Neighbor Distance (ND) and Patch Accuracy (PA) parameters used in the method were optimized. The Artificial Bee Colony Algorithm (ABC) was used for optimization. The proposed method is compared with the results obtained by the expert on test images in the problem of automatic counting of ovarian follicles. When the expert results were taken as a reference, the accuracy value of the method was found to be 96.75%. The method was also compared with the Faster R-CNN model. The Faster R-CNN model is accepted as a high accuracy model in object identification in literature. In this model, different CNN models can be used as layers. For this reason, AlexNet, Vgg16, and Vgg19 models are used as layer in Faster R-CNN. As a result of the experimental studies, it has been shown that the proposed method is more successful than the the Faster R-CNN model. Keywords: Deep Learning, Convolutional Neural Networks, Faster R-CNN, VggNet, AlexNet, Ovary, Follicle, Medical images, Artificial Bee Colony