1. Introduction
The incidence of kidney diseases has recently undergone considerable changes worldwide. Among these diseases, chronic kidney disease needs to be particularly researched1. It is noteworthy that most patients with kidney disease are early patients, having more treatment opportunities than those with advanced renal insufficiency. Regular screening allows patients to receive treatment in a timely manner to prevent and control kidney damage. The analysis and diagnosis of clinical urinary particles are important for screening and preventing kidney as well as related diseases2,3, because it can provide information about the type and quantities of cells in the urine, which, in turn, can provide a scientific basis for diagnosis by doctors. Therefore, it is crucial to accurately analyze and detect the types and quantities of cells in urinary particles.
Urinary particles are formed from the urinary tract and refer to substances formed by exudation, discharge, shedding, and concentration of crystals in visible form4. Common urine cells include erythrocytes, leukocytes, epithelial, casts, crystals, bacteria, fungi, etc.5,6. Among them, erythrocytes, casts, and crystals can be subdivided into several types. Depending on the type and number of cells, each component has a clear pathological significance. For example, erythrocytes in normal urine are invisible; however, when they appear in urine, it may indicate pathological bleeding in the urinary system. Common examination methods for urinary particles are microscopy7, SM staining8 and morphological examination. Among them, microscopy is the “gold standard” for urine analysis9 and can accurately detect cells in urine. The SM staining method requires staining of a urine smear, and the detection accuracy is further improved by using microscopy in combination with SM staining. However, the first two methods require manual operation, which is time consuming and labor intensive, and cannot meet the actual needs of clinical tests and the detection of automated urinary particles.
Urine morphological examination detects different cellular components by examining the size and shape of the cells under a microscope. According to different detection algorithms, the method can be roughly divided into two categories: traditional and new object detection algorithms based on artificial neural network. A typical object detection algorithm includes a detection algorithm based on decision tree classifier10, a support vector machine and template matching (SVM) 11, and an adaptive discrete wavelet entropy energy algorithm12. Traditional detection algorithms require the use of steps, such as object segmentation and manual design of feature extractors, and the speed and accuracy of detection need to be improved. With the rise and development of deep learning13, the object detection algorithm based on artificial neural network has achieved great success. It mainly includes a one-stage target detection algorithm represented by Yolo14,15 and SSD16, a two-stage target detection algorithm represented by Faster-RCNN17,18,19, and various variants derived from these algorithms20,21. Compared with traditional recognition algorithms, these algorithms have considerably improved the speed and accuracy of detection.
The object detection algorithm based on deep learning has been applied for the detection of urinary particles; however, the accuracy of detection needs to be further improved. In terms of the speed of detection22, the author pointed out that the slow detection speed was due to a class imbalance between the object class and the background class. In the online hard example mining algorithm23, the author increased the weight of the misclassified sample but ignored the easily categorized samples. Therefore, these algorithms have room for improvement in terms of detection speed. Most importantly, fewer cell types have been detected in previous studies due to three main reasons: the detection of multi-class cells is difficult, and it is a huge challenge for network design and parameter adjustment; as the number of categories increases, the confusion between cell categories increases, ultimately affecting the performance of the model; owing to the small size of cells and the large differences in the cell characteristics, character-stabilized, easily identifiable cells have been studied. However, the clinical significance of other categories is also huge and should be considered.
Based on the above discussion, we propose a method for analyzing and detecting urinary particles using multi-class fine classification based on deep learning. Compared with the previous urine object detection method, we have added many other rare but important categories of clinical significance that will not affect the accuracy and speed of detection, allowing quick and easy detection of multiple categories of urine cells and providing doctors with more abundant disease information, which can be of great significance for the clinical examination of urinary particles.