Legend

FIGURE 1a : The sample slide.
FIGURE 1b : Microscopic image acquisition system.
FIGURE 1c : The original image.
FIGURE 1d : Labeling work interface.
FIGURE 2 : The RetinaNet summarized into three parts: Resnet, FPN, and full convolutional neural (FCN) network. The FCN contains classification and regression subnets, as shown by the processes within the Class + box subnets.
FIGURE 3 : Feature pyramid network (FPN) includes bottom-up and top-down processes. C1–C5 represents convolution module, M2–M5 are the results of FPN up-sampling, and P2–P5 are feature map pyramids of object detection.
FIGURE 4a : Representative detection results of urinary particles using the detection models.
FIGURE 4b : Representative detection results of urinary particles using the detection models.
FIGURE 5a : The original image detection effect diagram.
FIGURE 5b : The image clarities of 6b is lower than 6a.The cell pointed by the red arrow represents leak recognition, and the cells indicated by the green arrows represent recognition errors.
FIGURE 5c : The image clarities of 6c is lower than 6b.The cell pointed by the red arrow represents leak recognition, and the cells indicated by the green arrows represent recognition errors.
FIGURE 5d : The image clarities of 6d is lower than 6c. The cell pointed by the red arrow represents leak recognition, and the cells indicated by the green arrows represent recognition errors.
FIGURE 6a : Cell stacking detection results. The red black arrows show the stacking of normal erythrocytes .
FIGURE 6b : Cell stacking test detection results. The black arrows show the stacking of oxalate crystals.