Fast classification and recognition method of blood cells using deep learning based on wrapped phase in polar coordinate

Abstract

In order to realize the nondestructive, fast and accurate classification of batch of blood cells, a classification method using deep learning based on wrapped phases in polar coordinate is proposed. In this method, after the coordinate transformation, the wrapped phases of cells are obtained first, and then three characteristic parameters of symmetry, roundness and singularity are established to analyze the phases. Subsequently, the data set of wrapped phases is built with the data augmentation. Finally, the classification and identification of batch of blood cells are realized through the convolutional neural network training based on LeNet-5 and the stochastic gradient descent algorithm (SGDM). An excellent result of the classification accuracy rate of 100% is obtained, which proves that the method has good feasibility and high accuracy. This method only requires the wrapped phases of cells in polar coordinates, so the complex unwrapping procedure and the additional noise effects are avoided.