Urinary Red Blood Cells Extraction Based on Multiple Instance Learning and Watershed Segmentation

Abstract

Doctors can diagnose many diseases by urinary red blood cells morphological analysis, and cell extraction is an important step in morphological analysis, which can help the subsequent classification of cells. However, it is time-consuming and inefficient for doctors to observe cells under a microscope, which easily leads to misdiagnosis and missed diagnosis. This paper proposes a method of cell extraction: firstly, the weak supervised multiple instance learning (MIL) is combined with resnet for cell localization, and the probability value of each pixel in urinary red blood cell image can be obtained through neural network, we set threshold value for probability value to obtain mask, finally, a marker-based watershed segmentation method was used to extract single urinary red blood cells. The experiment showed that the model achieved good results, with this method, doctors can directly analyze segmented cells, which can improve the efficiency of analysis and reduce the workload.