White Blood Cell Differential Counts Using Convolutional Neural Networks for Low Resolution Images

摘要

The Complete Blood Count (CBC) is a medical diagnostic test concerned with identifying and counting basic blood cells such as red blood cells (RBC), white blood cells (WBC) and platelets. The computerized automation of CBC has been a challenging problem in medical diagnostics. In this work we describe a subcomponent system for the CBC to perform the automatic classification of WBC cells into one of five WBC types in low resolution cytological images. We describe feature extraction and consider three classifiers: a support vector machine (SVM) using standard intensity and histogram features, an SVM with features extracted by a kernel principal component analysis of the intensity and histogram features, and a convolutional neural network (CNN) which takes the entire image as input. The proposed classifiers were compared through experiments conducted on low resolution cytological images of normal blood smears. The best results were obtained with the CNN solution with recognition rates either higher or comparable to the SVM-based classifiers for all five types of WBCs.