Red Blood Cell Image Classification Using Model Observers

Healthy red blood cells (RBCs) undergo a gradual morphological transformation during storage. From original healthy discocytes, RBCs gradually lose membrane surface area and cell volume, eventually turning into ghost (lysed cell). The degree of deterioration varies from cell to cell. These cells can be classified into 7 classes: discocyte, stomatocyte, echinocyte 1, echinocyte 2, echinocyte 3, sphero-echinocyte and spherocyte. Currently, researchers categorize a blood cell into different class by visual inspection. This work is laborious and inefficient. The limitation on the sample size is a problem for the evaluation of the blood unit quality and other research. Our objective was to test linear discriminants for classification work. The images we used were provided by Nate and his colleagues who fabricated a microfluidic device that can take image of RBCs. We extract features based on cell shape and surface texture. The overall accuracy of our system is 69.4%.