4. Discussion
Based on the RetinaNet model technology of deep learning, we established a multi-class classification method for the detection of urinary particles. The average accuracy of recognition was 82.86%, and the detection accuracy of a single category was also high. Compared with other methods, our method was able to perform multi-class analysis and detect urine cells. The experimental results showed that the accuracy of recognition is relatively high, with some categories showing better accuracy than others; good results were also achieved in terms of speed performance. However, using experimental data, we found that the recognition effect of some categories was unsatisfactory. Therefore, we discuss the specific factors that affect the accuracy of recognition result through both internal factors and external interference.
In the early stage of model design, the network structure and network parameters are often determined using empirical values. Different network parameter configurations will have different effects on the network model. As shown in Table 2, we discussed the impact of the following four conditions on accuracy: weight initialization method, feature extractor selection, anchor size, and loss function parameter configuration.
The experimental results demonstrated that the accuracy of the model was higher in the COCO weight initialization mode. The experimental results of Resnet50 and Resnet101 basic networks were very close; however, considering the Resnet101 network model is deep and more complex, we chose the Resnet50 basic network. In addition, anchors with a smaller size can help in improving the accuracy of the method. In the parameter selection of loss function, we observed that when αt was 0.25 and γ was 2.0, the effect was better.
Based on the above discussion, we chose a satisfactory model parameter configuration. In addition, medical images are expensive to label, and a lack of pathological samples could result in a lack of sample data in certain categories. However, deep learning is based on multi-dimensional data extraction and analysis of big data. The lack of data could fundamentally affect the accuracy of model testing. In this study, due to insufficient data regarding uric acid crystals, low-transitional epithelium, and abnormal erythrocyte, the recognition accuracies of these types are low, thus affecting the accuracy of the method.
In order to compare with other methods, we compare the optimized model results with two other typical methods. The comparison results are shown in Table 3. The results show that our accuracy rate is higher than the other two methods. This method also has advantages in processing a single image. Therefore, this method is very helpful for the clinical diagnosis and automated detection of urinary particles.
When the focus of the objective lens is not clear, poor image quality could lead to recognition errors or missing recognition. In order to explore this influencing factor, we artificially changed the focus of the objective lens and photographed four sets of images using different sharpness in the same field of view. The acquired image was input into the model for object detection. The detection effect diagram is shown in Figure 5a, Figure 5b, Figure 5c and Figure 5d. As observed in the figure, as the degree of blurring of the image deepens, the situation regarding the leak recognition becomes more serious. Compared with Figure 5a, there is a significant difference in the sharpness of the image, and many category recognition errors appear in Figure 5d. Therefore, beyond a certain range, the quality of the image can also affect the accuracy of recognition.
Similar to focus blur, when the shape of the cell changes (during cell degradation) the data characteristics of the cell change, affecting the accuracy of identification. Therefore, we should conduct timely processing of the detected sample or perform human interference (such as refrigeration of the sample) to prevent degradation.
Due to the characteristics of easy adhesion between cells, the phenomena of overlap, stacking, and even agglomeration occur, causing leakage recognition of the model and affecting the detection results of the model. As shown in Figure 6a, Figure 6b, due to the stacking of cells, the recognition algorithm treated it as a single cell, whereas others were not recognized. This is a flaw of the method. Thus, the algorithm has more room for improvement and research.