Segmentation of microscopic cell images using deep learning methods

The cell is the smallest structural unit essential for organisms to perform fundamental biological functions. Cell analysis plays a vital role in both basic biological research and the clinical diagnosis and treatment process. This analysis is a complex process that requires the interpretation of medical data. In this process, cell classification and cell segmentation techniques are used to make sense of medical images. Cell classification involves methods that differentiate cells based on specific biological characteristics. Cell segmentation involves techniques that focus on detecting the boundaries of individual cells. These techniques face several challenges, such as the variable nature of microscopic cell images, different staining and imaging techniques. Various deep network architectures have been developed to overcome these challenges. In this thesis, applications for cell segmentation and cell classification tasks are developed based on deep network architectures. First, the cell segmentation performance of different deep network architectures are analyzed. In the study, The DeepLabv3+ architecture achieved the best results, with a Jaccard index of 0.8957 and a Dice coefficient of. In subsequent studies, a new deep learning architecture, RaUNet, was proposed to perform the cell segmentation task. The RaUNet architecture was created by integrating the inverse attention module into the skip connections of the U-Net architecture. The inverse attention module reduces the information loss in the U-Net architecture and preserves detailed information. RaUNet architecture is tested on four different datasets. Third, we focus on the input image quality in the cell segmentation task. In this application, cell images are preprocessed using wavelet transform and images with different wavelets are combined to prepare for the segmentation task. The application was tested on two datasets. Finally, a system that classifies white blood cells after segmentation is designed. This system utilizes of the features obtained by various deep network architectures in the transfer learning process and an effective feature selection process. The proposed system showed the highest classification performance with an accuracy of 0.9863 on the BCCD dataset. The applications developed within the scope of this thesis have a remarkable potential to strengthen the decision support systems of physicians thanks to the innovative and unique approaches they offer compared to current methods.