Detection and classification of cells in urine sediment images with deep learning method

In today's world, one of the most requested tests by doctors is the urine test. One of the main reasons for this is that urine contains a wealth of information about human metabolism. Another reason is that urine sample collection is relatively easy. Fully automated urine analyzers are used for the analysis of the collected samples. However, many of these analyzers do not have the ability to learn retrospectively. In other words, it is difficult to teach the device about a substance that has not been previously introduced. During the analysis, these devices typically utilize primitive image processing algorithms such as edge and color detection. These algorithms can lead to the inclusion of physical objects such as air bubbles in the measurement. This poses a significant problem for laboratory technicians in terms of the accuracy of the test. In this thesis, the problems in urine analyzers mentioned above were tried to be solved by deep learning method. Three different datasets, two of which are ready and one from the urine analysis device, were obtained and some experiments were performed on these datasets. As a result of these experiments, an accuracy rate of 94% in erythrocyte detection and 87% in leukocyte detection was achieved. Thus, particles in human urine were successfully detected using the YOLOv7-tiny deep learning model.