Classificação de Leucemias Utilizando Aumento de Dados, Transferência de Aprendizado e Combinação de CNNs

Abstract

Early diagnosis of leukemia significantly increases patients’ chances of cure. But for this, it is necessary to be aware of which primary form the patient has since it can manifest in two ways: acute and chronic; these are still subdivided into myeloid and lymphoid. The initial focus of the present study is the acute form, a particular type of leukemia that causes abnormal cell growth in a short period, requiring an accurate and rapid diagnosis to increase the chances of successful treatment. In sequence, we expanded the focus of the work to the classification in other types of leukemia. Deep learning models are an ally in these diagnoses, which have been increasingly used in computer-aided medical diagnostic systems developed to detect leukemia. This study proposes a Convolutional Neural Network - CNN) called Acute Leukemias Recognition Network (AlertNet) that was inspired by convolutional blocks of the Vgg16 network but with smaller dense layers. To define the parameters of the AlertNet network and its variations, we evaluated different models of CNNs and fine-tuning methods using eighteen datasets of images with different resolution, contrast, color, and texture characteristics, which totaled 3,536 images. In addition to this network, we also used seven pre-trained networks to evaluate and compare results. We applied data augmentation operations to expand the training dataset and divided it into five leukemia classification scenarios: three binary and two multiclass classification problems. To assess the generalizability of CNNs, we applied a cross-validation technique to the dataset. The results of the experiments were promising, with 96.17% accuracy for the classification of acute leukemia and with 94.73% and 94.59% of correct answers obtained by the multilevel and committee in a scenario with four classes, respectively. Methods such as multilevel and ensemble were applied in this study to improve networks’ performance. These methods have contributed to reducing the prediction error and variance, improving the model’s accuracy.