An artificial intelligent diagnostic system on differential recognition of hematopoietic cells from microscopic images.

Abstract

Despite their advantages, none of the automated white blood cell differentiated counters have replaced the conventional microscopic evaluations of blood and bone marrow slides by hematologists. We have analyzed the smears of 39 patients and 8 control subjects to develop an artificial expert system that recognizes 16 different types of nucleated hematopoietic cells during the stages of differentiation. A charge coupled television camera and a special frame grabber were used for data acquisition, and 247 nucleated cell images were transferred from a microscope to an IBM 386 computer to be processed. One hundred sixty-five and 82 of these images were used for training and testing, respectively. Our system is composed of image processing and analysis (enhancement, thresholding/smoothing, edge detection), pattern recognition (feature extraction and classification with supervised artificial neural network), and expert system development. Image processing and analysis were used to obtain 13 cellular features to be used as the input parameters (neurons) of the artificial neural network. A supervised artificial neural network (back-propagation learning algorithm) was used in the classification of 16 different cells (output neurons of the neural network), which is the second step of pattern recognition. A confusion matrix has been developed to compare the similarities and dissimilarities between the differential recognitions of the hematologist and the expert system. The discriminatory power of the procedure is statistically significant: Q = (N - n.K)2/N.(K - 1) = 28.2. The sensitivity and the specificity of the expert system were 71.4% and 90.9%, respectively.