Leukaemia is a dysfunction that affects the production of white blood cells in the bone marrow. Young cells are abnormally produced, replacing normal blood cells. Consequently, the person suffers problems in transporting oxygen and in fighting infections. This article proposes a convolutional neural network (CNN) named LeukNet that was inspired on convolutional blocks of VGG-16, but with smaller dense layers. To define the LeukNet parameters, we evaluated different CNNs models and fine-tuning methods using 18 image datasets, with different resolution, contrast, colour and texture characteristics. We applied data augmentation operations to expand the training dataset, and the 5-fold cross-validation led to an accuracy of 98.61%. To evaluate the CNNs generalisation ability, we applied a cross-dataset validation technique. The obtained accuracies using cross-dataset experiments on three datasets were 97.04, 82.46 and 70.24%, which overcome the accuracies obtained by current state-of-the-art methods. We conclude that using the most common and deepest CNNs may not be the best choice for applications where the images to be classified differ from those used in pre-training. Additionally, the adopted cross-dataset validation approach proved to be an excellent choice to evaluate the generalisation capability of a model, as it considers the model performance on unseen data, which is paramount for CAD systems.

Leukemia is a disorder that affects the bone marrow, causing uncontrolled production of leukocytes, impairing the transport of oxygen and causing blood coagulation problems. In this article, we propose a new computational tool, named LeukNet, a Convolutional Neural Network (CNN) architecture based on the VGG-16 convolutional blocks, to facilitate the leukemia diagnosis from blood smear images. We evaluated different architectures and fine-tuning methods using 18 datasets containing 3536 images with distinct characteristics of color, texture, contrast, and resolution. Additionally, data augmentation operations were applied to increase the training set by up to 20 times. The k-fold cross-validation (k = 5) results achieved 98.28% of accuracy. A cross-dataset validation technique, named LeaveOne-Dataset-Out Cross-Validation (LODOCV), is also proposed to evaluate the developed model’s generalization capability. The accuracy of using LODOCV on the ALL-IDB 1, ALL-IDB 2, and UFG datasets was 97.04%, 82.46%, and 70.24%, respectively, overcoming the current state-of-the-art results and offering new guidelines for image-based computer-aided diagnosis (CAD) systems in this area.