Bone Marrow Cell Recognition Technology Based On Multi-classifier Fusion Research

The computer-aided automatic classification of bone marrow cell images is of great importance for the clinical diagnosis of many fatal diseases. However, due to the hematopoiesis in bone marrow, the distribution of cell types are very complex so that the traditional recognition methods for the peripheral blood cells herein can hardly attain satisfying result. This thesis tries to apply new techniques of multiple classifier fusion to the area of identification of the marrow cells, and make efforts to meliorate the fusion models.Multiple classifier fusion, or combination, is a modern technique in pattern recognition areas. Through pertinently combining different information from varies of simple classifiers, the classification accuracy can be fairly improved and the difficulty of designing a single, high-accuracy classifier could be avoided. In recent years, fusion methods of many kinds have been widely used in the identification of human face, hand-written characters, remote sense images, etc., but relatively rarely studied in the medical image region.The construction of a fusion system mainly concerns to three steps: Devising individual classifiers, selection of the component classifiers, and designing proper fusion model to combine these components. In this paper, after cell features selected by genetic algorithms and BP networks trained as individual classifiers, a dynamic classifiers selection method, which integrates the minimization of correlation and the maximization of reliability, is proposed to get the optimized component classifiers. By analyzing the different fusion theories and comparing their performances, meliorated knowledge-based fusion models are addressed using D-S correlative evidence theory and adaptive fuzzy integral so that the variable capability of the component classifiers according to the input testing samples could be considered during the combination. Experiments on marrow cells testified its superiority with traditional classification methods both in accuracy and feasibility.