Simultaneous localization and classification of acute lymphoblastic leukemic cells in peripheral blood smears using a deep convolutional network with average pooling layer

Abstract

It is important to analyze and classify the blood cells for the evaluation and diagnosis of many diseases. Acute Lymphoblastic Leukemia (ALL) is a blood cancer mostly found in children below the age of 7–8 years. It can be fatal if left untreated. ALL cells are abnormal lymphocytes that have a condensed appearance to their chromatin. ALL can be detected through the analysis of white blood cells (WBCs) also called as leukocytes. Presently the morphological analysis of blood cells is performed manually by skilled operators, which makes it a time-taking and non-standardized process. This paper presents a novel deep learning approach to automate the process of detecting ALL from whole-slide blood smear images. Previous work in this domain deal with the isolation and classification of WBCs based on certain morphological image features. However, this work uses a deep network for simultaneous localization and classification of WBCs. The network makes use of Average Pooling layers to figure out the hot-spots of WBC locations in whole-slide images. Although this workflow fails to successfully figure out all the ALL lymphocytes in a whole-slide image, but it does performs very well in the task of predicting whether the blood smear image belongs to an ALL patient or not.