Automatic Classification of Erythrocytes Using Artificial Neural Networks and Integral Geometry-Based Functions

The red blood cell deformation caused by disorders like sickle cell disease can be assessed by observing blood samples under a microscope. This manual process is cumbersome and prone to errors but can be supported by automated techniques that allow red blood cells to be classified according to the shape they present. There are proposals in the literature that use functions based on integral geometry to obtain a description of the cells' contour before performing classification, reaching 96.16% accuracy with the use of the k-Nearest Neighbor (KNN) classifier. In those approaches, the classification-confusion cases persist mainly in the classes of most significant interest, which are those related to the detection of deformed cells. In this work, we use artificial neural networks-based classifiers, trained with the characteristics obtained from integral geometry-based functions, to classify erythrocytes into normal, sickle, and other deformations classes. Our proposal achieves accuracy of 98.40%. This result is superior to those of previous studies concerning the classes of greatest interest. Also, our approach is computationally more efficient than previous works, making it suitable for supporting medical follow-up diagnosis of sickle cell disease.