Abstract Aim: Traditional artificial microscopic technologies cannot
meet the current demands of automated urine detection. Furthermore, the
number of cell types detected in previous studies was relatively
limited; therefore, previous studies are considered to be insufficient.
Methods: The present study proposes a multi-class detection method of
urinary particles based on deep learning. First, we obtained an image
database containing 15 types of cellular components, i.e., normal,
shrinking, glomerular, and abnormal erythrocytes; leukocytes; calcium
oxalate, uric acid, other types of crystals; particle and transparent
casts; epithelial cells; low-transitional epithelium; Candida; Bacillus;
and abnormal epithelium. The image data was then input into Resnet50
basic network and feature pyramid network (FPN) to obtain a multi-layer
feature map. Thereafter, the classification sub-networks and regression
sub-networks were used to classify and locate the cellular components.
The network detection model was obtained after training was completed.
Results: The experimental data showed that for the test set, the mean
average precision (mAP) of the network model reached 82.86%, and the
time required to process a single image sample was 195 ms. Therefore, we
were able to perform multi-class analysis and detect urine cells with
good results in terms of detection speed. Conclusion: This study applies
the deep learning network model for the multi-category detection of
urine cells. The method can be used to analyze and detect urinary
particles in actual clinical practice and has great reference
significance for the detection of other cells in the clinic.