An Improved YOLOX for Detection in Urine Sediment Images

In clinical medicine, the detection of human urine sediment is usually a basic test. The indicators of test can effectively analyze whether the patient has the disease. The traditional method of chemical, physical and microscopic analysis of samples artificially is time-consuming and inefficient. With the development of deep learning technology, many detectors now can replace traditional manual work and play a good detection effect. So it is of great value to put deep learning technology into the medical field. Usually, the Urine Microscopic Image has challenges for research, in which many detectors can not detect the cells well due to their small scale and heavy overlap occlusion. Therefore, we propose a novel detector for the detection of urine sediment in this paper. Firstly, considering the friendliness of YOLOX to the small objects, we adopt the framework from the YOLOX. Secondly, we add spatial, channel and position attention to enhance the feature information to achieve more accurate detection results. Then, the better Giouloss is also applied to make a better regression of the bounding box. Finally, the experimental results show that our improved model based on YOLOX achieves 44.5% AP(50-90 )and 80.1% AP(50) on the public dataset Urine Microscopic Image Dataset, which is far better than other detectors.