Determining the microscopic content of human urine sediment by deep learning based artificial intelligence system

Urine is examined with a microscope; it can give important information about the body health. At the hospitals, a lot of urinary analysis are done on automatic urine analyzers and medical information is obtained for diagnosis or treatment. It is thought that urine analyzers, which use traditional image processing techniques, can be further developed if deep learning based artificial intelligence methods are used in image processing unit. In this study, the contents of bacteria, epithelium, erythrocyte, leukocyte, crystal, and yeast in microscopic images of human urine were determined by implementing Mask R-CNN deep learning system. It can detect objects in images and do instance segmentation. ResNet101 and ResNet50 CNN models are used as backbones. In order to create a Mask R-CNN model that can operate with less system resources, the MobileNet network has also been adapted and worked with a total of three backbones. Also, examine the effect of size difference on object detection, different values of the RPN anchor scale parameter were evaluated. In object detection with Mask R-CNN, two types of boundaries are drawn as masks and bounding boxes. For this reason, separate performance evaluations were made for masks and bounding boxes. A test set with 1154 urine content patterns was used to test the system. The highest detection rate of these patterns was 69.15% for masks and 73.31% for bounding boxes when the MobileNet backbone was used. The best mAP values were obtained using the ResNet101 backbone; 0.774 for masks and 0.836 for bounding boxes. These performance values can be developed training with more urine images and better optimization. With this study, it has been shown that Mask R-CNN deep learning system can be used for urinalysis.