Automatic zone identification in blood smear images using optimal set of features

Visual assessment of peripheral blood smears is an important diagnostic approach in the hematology. The first step in such analysis is to identify the appropriate regions on the slide for screening. However, observing numerous samples of blood slides under the microscope by a hematologist is a very slow, inconsistent and exhausting job that raises the error possibility. Digital microscopes with the help of image processing techniques can do this procedure automatically. We proposed an algorithm to automatically classify smear images into “Good”, “Clumped” or “Sparse” regions. We first segment the cells using an adaptive thresholding technique and then extract their central zone using two different approaches, then a total of 24 features is extracted from the segmented regions, three of them are newly introduced to better quantify the cell spreading and clumping. Unlike the other studies, to elevate the classification results we select an optimal subset of features through feature selection experiments. The experimental results on 2400 blood smear images show average classification accuracy of 98.5%. Also, sensitivity and specificity for finding “Good” working areas are gained to 97.6% and 99.0%, respectively. In comparison with the most state-of-the-art algorithms, our approach improves the evaluation measures and computation time dramatically.