White Blood Cell Classification Based on Shape and Deep Features

Classification of the white blood cells (WBCs) in blood smear images is essential for providing important information to the physicians. In addition, manual analyzing of the blood smear images for determining the various WBCs is a time-consuming issue for the physicians. In this paper, a hybrid method is proposed for the classification of the WBCs. Image processing (IP) and machine learning (ML) are used to determine and classify the WBCs in blood smear images. In the IP perspective, various IP algorithms are used to segment the WBCs and in ML perspective, feature extraction and classification are employed. RGB to HSV transformation, color to gray tone conversion, filtering operations, thresholding, and morphological processes are used for determining the WBCs. Median filtering and adaptive histogram equalization are used for filtering and image enhancement and Otsu thresholding is considered for thresholding due to its simplicity. Shape based features and deep features are used for characterization of the WBCs and long-short term memory (LSTM) network is employed for classification. A dataset containing totally 349 blood smear images is considered in the evaluation of the proposed method. 10-fold cross validation is used in experiments and classification accuracy is calculated accordingly. While the shape based features produce 80.0% accuracy, deep features obtain 82.9% accuracy. When both shape and deep features are concatenated, 85.7% accuracy score is obtained.