Classification of red blood cell images using a neural network.

Blood smears from four persons were photographed at 400x magnification; one person had normal cells (NC), one had microcytic hypochromic (MH) anemia, one had sickle cell (SC) anemia and one had hereditary spherocytosis (HS). The photomicrographs were transferred to computer media as 8-bit greyscale image files using a slide digitizer. The images were further manipulated using an image processing program, (NIH Image). Corrections were made for variation in background intensity of the images. Individual RBC were manually cropped from the photomicrograph images into separate files of 75 x 75 pixel size. From each RBC image, 5 values the area, mean pixel density, standard deviation of pixel densities, integrated density and modal density were obtained for pixels with greyscale values above the image noise threshold; two thresholds were used. These parameters were chosen for ease of acquisition and because they appeared to be independent descriptors of RBC morphology. Each set of 5 values constituted a pattern for presentation to the artificial neural network. At each threshold, 61 RBC were measured, with 14 NC, 16 MH, 13 SC and 18 HS cells, resulting in 61 patterns. For training the ANN, patterns at the same noise threshold were presented, as far as possible, in the order of one pattern per category following the other. As an important data pre-processing step prior to training and testing the ANN, all 61 values for each parameter were normalized to lie between -1 and +1.