Image analysis of nucleated red blood cells

Bone marrow smears stained with Giemsa were scanned with a video camera under computer control. Forty-two cells representing the six differentiation classes of the red bone marrow were sampled. Each cell was digitized into 70 X 70 pixels, each pixel representing a square area of 0.4 micron2 in the original image. The pixel gray values ranged between 0 and 255. Zero stood for white, 255 represented black, while the numbers in between stood for the various shades of gray. After separation and smoothing the images were processed with a Sobel operator outlining the points of steepest gray level change in the cell. These points constitute a closed curve denominated as inner cell boundary, separating the cell into an inner and an outer region. Two types of features were extracted from each cell: form features, e.g., area and length, and gray level features. Twenty-two features were tested for their discriminative merit. After selecting 16, the discriminant analysis program classified correctly all 42 cells into the 6 classes.