Blood Cell Image Classification Based on Image Segmentation Preprocessing and CapsNet Network Model

The identification and examination of peripheral white blood cells can help Department of Hematology doctors diagnose AIDS, leukemia and blood cancer and other diseases. Experts' classification of white blood cells in the blood is a very complicated and time-consuming task. Subjective factors such as human experience and even fatigue can have a great impact on the accuracy of recognition. The computer image processing system can automatically complete the task of medical image analysis to shorten the analysis time of the doctor, eliminate the influence of subjective factors, and finally improve the accuracy of the recognition. Our software framework uses a deep learning model, the CapsNet network model, to analyze white blood cell images of peripheral blood smears. Four major types of white blood cells (eosinophils, lymphocytes, monocytes, neutrophils) can be classified. First, the U-Net convolution neural network is used to segment the white blood cell image, and then the CapsNet network model is established to classify and predict the segmented white blood cells images. After the experiment, we obtained a lot of line graphs including precision rate, margin loss, reconstruction loss and total loss using the tensorboard tool, and the test results achieved good results. The final classification results are as follows: the classification accuracy of the white blood cells image test set is 85% and the classification accuracy of the train set is 99%. The classification accuracy of our experimental methods is significantly higher than traditional machine learning, such as 2.7% higher than Bayes classifier and 14.4% higher than k-Nearest Neighbor (KNN).