Rede Neural Convolucional para o Diagnóstico de Leucemia

Leukemia is a type of cancer that affects the production of blood cells in the bone marrow which makes it challenging to coagulate blood and fight infection. In this work, we propose a method for the automatic diagnosis of leukemia using Convolutional Neural Networks (CNNs). We use pre-trained CNNs and learning transfer techniques in constructing the proposed method. We employed the Deeply Fine Tuning Modified (DFTM) technique combined with data augmentation operations to fine-tune the pre-trained model. To train and test the proposed method, we used a set of 2304 images from 14 different image databases. The proposed method reached an accuracy of 98.84%, and when compared to other works, we observed greater robustness and consistency in the results. We conclude that the fine-tuning is more robust the classification of heterogeneous images when compared to the features extraction through CNNs.