3. Results
After 50 epoch trainings and a total of 62 hours, we developed a fine
classification model having an accuracy of 86.67% on the training set.
The accuracy of the test set reached 82.86%, and a good prediction
result was obtained. Figure 4a, Figure 4b shows the recognition effect
of the image of the urinary particle after the model test. Table 1 lists
the average accuracy results for each category of the model. Among them,
the crystals class had the highest accuracy of 92.26%, whereas those of
abnormal erythrocytes, abnormal epithelial cells, and calcium oxalate
crystals were not ideal. We discuss the specific reasons behind this
result in the next section.