Obtaining Deep Learning Models for Automatic Classification of Leukocytes

Introduction

The number of leukocytes presdataent in the blood, also known as white blood cells, provides important information regarding the state of the immune system, allowing to evaluate potential health risks. A significant change in the number of leukocytes, relative to reference values, is usually a sign that the body is affected by some type of antigen. Moreover, the variation in a particular white blood cell type is generally correlated with a specific type of antigen.

White blood cells are generally classified into 5 categories: lymphocytes, monocytes, neutrophils, eosinophils, and basophils. There is also the band designation for a specific form of the nucleus. Figure 1 shows examples of these categories.

Figure 1.

Leukocyte types (from left to right): Neutrophil, Eosinophil, Basophil, Lymphocyte, and Monocyte (Noble, 2019)

The preparation and analysis of blood samples are usually affected by deviations naturally introduced by manual operations. These difficulties can be minimized when performed by highly trained technicians. However, these tasks are labor-intensive and time-consuming and always subject to error. For these reasons, there is interest in having systems that can automatically classify with high specificity and high sensitivity.

The evolution of techniques for counting and identifying leukocytes (and blood cells in general) began in the mid-nineteenth century with the use of capillaries and slides. Over the years, several types of devices designed to count blood cells have appeared, which would later enable their classification.

The classification of blood cells has always been done by human specialists until the 1960's, since when emerged the possibility of automating this task. Firstly, through the use of optical and impedance methods and later through algorithms developed specifically for this purpose from microscopy images within the scope of computer vision and, in the last two decades, using neural networks.

Several researchers have presented papers for leukocyte identification and counting. Techniques that use classic machine learning models, in opposite to deep learning models, are built on manually selected characteristics. This approach can use shallow neuronal networks (Rezatofighi & Soltanian-Zadeh, 2011; Nazlibilek et al., 2014), Support Vector machines (SVM) (Rezatofighi & Soltanian-Zadeh, 2011; Putzu et al., 2014; Agaian et al., 2014; Alférez et al., 2016; MoradiAmin et al., 2016), Bayes classifier (Stathonikos et al., 2013; Prinyakupt & Pluempitiwiriyawej, 2015), etc. The manipulation of characteristics, prior to the classification model, may involve segmentation (Chaira, 2014), extraction and selection of features that describe the leukocyte defining region (Alférez et al., 2016). Thus, we can divide these classic processes into three main stages: segmentation, feature extraction, and classification. These approaches have the advantage of allowing the use of relatively small datasets, as the segmentation and the features used reduce the variability of the patterns delivered to the classification models. On the other hand, the segmentation performance and the lack of universality of the descriptors can limit the result achieved by the classification models. An example of this approach is found in Dan et al. (López-Puigdollers et al., 2020), where features such as SIFT (Lowe, 2004) are employed. However, the assertiveness of the classification is not very high.

The use of deep learning models tends to solve the problems presented in the approach described above, provided that the dataset is sufficiently representative of the pattern variability, associated with leukocyte optical visualization, or a transfer learning approach is employed, as is the case of the present work.