An accurate classification method based on multi-focus videos and deep learning for urinary red blood cell
Xinyu Li, Ming Li, Yongfei Wu, Xiaoshuang Zhou, Hao Fang, Xueyu Liu · 2020
Generally, there are multitudinous red blood cells in the urine of patients with urinary system diseases (USDs). Classifying urinary red blood cells (U-RBCs) is an important diagnostic method for USDs. Because the U-RBCs classification relying on professional microscopist is difficult to satisfy the growing urinalysis demand of patients, researchers have proposed several U-RBC classification methods based on artificial intelligence technologies. However, the existing methods mainly using single-focus images are unable to accurately classify U-RBCs because the shapes of U-RBCs may deform significantly with the change of microscope focus. Therefore, the research in this paper aims to develop a high-accuracy U-RBCs classification method by the combination of innovative multi-focus videos and mature Inflate 3D Convnet model for video classification. Experimental results show that the proposed method achieves overall accuracy of 82.1% in classifying six categories of U-RBCs, which is obviously higher than traditional method with overall accuracy of 75.9%. Comprehensively, the proposed method is proved to be successful and significant for its advantages of accuracy, robustness and low cost of data collection. In the future works, this method can also be extended to accurately classify other medical targets under the microscope, and its classification performance would be further improved by introducing the multi-instance learning model.