Automatic Identification Technology Based On Multi-classifier Fusion Of Bone Marrow Cells

Bone marrow is the main hematopoietic organ in clinical medicine. Alterations in the morphology, number and species of bone marrow cells commonly signal an internal disease. Relative to traditional classification methods, using computer-aided diagnosis to classify the cells in the images of bone marrow is more efficient and faster, which is of great significance for auxiliary diagnosis. The study of automatic recognition of bone marrow cells is particularly challenging. The features and categories of the bone marrow cells are complex. Moreover, although the automatic recognition of peripheral blood cells is mature, it’s not satisfactory at classifying bone marrow cells. In this paper, the multiple classifier fusion method is applied to the classification of bone marrow cells. The main works are listed as follows:A gray gradient based Hough transform algorithm is proposed to detect and segment bone marrow cells. The gradient based Hough transform algorithm finds coordinates of the center of a cell by using a two-dimensional array containing the accumulated value. This algorithm is efficient and speedy. Experiments show that this algorithm is adaptive to both normal and abnormal cells.Thirty-four features of bone marrow cells are selected, which are categorized in three groups: morphology, color and texture. The feature weights are analyzed and sorted by the Relief algorithm. Sixteen features are selected from the initial thirty-four ones. By reducing the dimension of features, the redundant features are also reduced. Besides, the design of single classifiers is simplified. The recognition efficiency is improved, thus reducing the execution time.Three single classifiers are designed, including the SVM classifier, the KNN classifier and the BPN classifier. The average recognition rates are61.9%,59.32%and86.23%respectively. It shows that the single classifiers are insufficient to the classification of bone marrow cells and it’s necessary to fuse multiple classifiers for increasing recognition rate.The outputs of single classifiers are fused using voting method, Bayesian method, decision template and evidence theory. The recognition rates are91.08%,91.77%,89.78%and91.86%respectively, which are better than those of single classifiers. The evidence weight is proposed, which improves the conflicting evidence of evidence theory. The weights of evidences are reassigned by calculating the distance between evidences and the credibility of the evidences. The improved evidence theory obtains an average recognition rate of94.62%, which is2.76%higher than the classic evidence theory and shows its superiority to other fusion methods.