Classificação de linfócitos utilizando imagens microscópicas de células de pacientes com Leucemia Linfoide Aguda

Abstract

Counting and classifying cells allows the diagnosis of countless diseases of the blood tissue, among of them Acute Lymphoid Leukemia (ALL), a cancer that affects lymphocytes and can be fatal if not correctly treated in early stages. The classification of lymphocytes into healthy or malignant is usually performed by trained pathologists from observation of microscopic images of patient’s lymphocytes. They look for for cytological and morphological features that allow a correct classification. Although there are advanced diagnostic methods, such as Flow Cytometry, they are still expensive and are not available in all hospitals and treatment clinics, this fact makes the microscopic analysis still one of most commonly used methods. However this analysis is time consuming, tedious and repetitive for the professional who perform it, besides producing non-uniform results and with low reproducibility. In this sense, the use of image processing techniques in combination with classification algorithms can help medical professionals by providing information on the cellular morphology of lymphocytes in a more standardized way and with more accurate results, besides allowing the creation of automated classification systems capable of differentiating between healthy and malignants lymphocytes. This work aims to apply image processing techniques and classification methods to lymphocyte images extracted from ALL patients in order to classify them as healthy or malignant. For this purpose the following image descriptors were extracted: First Order Statistics, Texture Features, Morphological Featues, Centroid Distance Function and Discrete Transform of Cosine coefficients. Each set of descriptors was used with each of the following classification algorithm: Support Vector Machine, K-Nearest Neighbors, Random Forest and Neural Networks. Also was trained the convolutional neural networks: InceptionV3, Xception, VGGNet and ResNet50 using raw images. Another goal of this work was to study how patient’s lymphocyte particularity affect the performance of classifiers. The dataset consists in 10,661 images, 7272 of which are healthy and 3389 malignant lymphocytes, which were extracted from 47 with ALL and 26 patients, respectively. The effects of the data augmentation technique when applied to the images were evaluated.