Optimal decision fusion-based systems of neural network ensembles for white blood cell classification

The classification and counting of white blood cells from samples taken from patients is an important activity in accurately establishing the diagnosis of certain diseases. The methods based on artificial neural networks for this task have gained considerable attention as decision-support systems in medicine. As a novelty, the authors have implemented a global white blood cell classification system (eight classes) using the optimal fusion of individual decisions from several performing networks. Three neural network ensemble systems were implemented using different network selection and decision fusion methodologies, starting from initial neural networks trained and validated on a publicly available database. In total, nine experimentally chosen convolutional neural networks were considered as individual classifiers. The systems were based on optimizing classification scores, which depended on each neural network and each blood cell class. Two systems used weights associated with individual networks calculated from the F1 score because it is a model evaluation metric that provides a better measure of both correctly and incorrectly classified cases. Another system used the probability of the input image belonging to the classes considered. The originality of this work lies in the methodology of selecting neural networks, the technique of combining them based on three approaches, and the decision-making procedure in the proposed collective systems. The benefits of the ensemble-based system include robustness, reduced risk of overfitting, and model generalization. The performance achieved by the systems was superior to individual networks or similar work, which can lead to progress in the medical field.