Study On Leukocyte Image Recognition Method Based On Mixed-attribute Fusion In Varied-space

Differential leukocyte count is of great clinical significance, and leukocyteexamination using automated hematology analyzers has been a major diagnosis methodfor doctors. Automated hematology analyzers based on image processing have not beenwidely applied for lack of mature and stable automated identification method forleukocyte images. Leukocyte image recognition method is of far-reaching researchsignificance and application prospect becouse of its major advantages: low cost andpopularization convenience. Focusing on mixed-attribute data classification andaccorrding on formal concept analysis theory, this research proposed a leukocyte imagerecognition method based on mixed-attribute fusion in varied-space. By making full use ofmixed-attribute information in varied-space, to solve the problems of technologyinstability and medical background complexity.This method not only exhibited theoreticalimportances for multiple attribute information fusion and pattern classification, but alsoprovided practical technique improves for automated hematology analyzers.Firstly, to extract attribute features of leukocyte images accurately, this researchproposed a robust methoed for leukocyte image segmentation, which included4steps:leukocyte locating based on the algorithm of component combination in multi-colourspace, adhesive leukocyte images segmentation based on the argument function algorithm,nucleus segmentation based on fingerprint histogram smoothing algorithm, and cytoplasmsegmentation based on overlap and subtraction algorithm. The experiment results showedthat, this method was able to locate leukocyte accurately and segment nucleus andcytoplasm perfectly in irradiant, tinctorial and adhesive microscopic images.Secondly, basing on formal concept analysis and attribute partial-ordered structuretheory, this research proposed a attribute-reduction and rule-extraction method forleukocyte images. The bottom features of shape, texture and color were extracted firstly,and then fewer image features were preselected by inert-class overlap coefficient matrix.After that, attribute partial-ordered structure was created based on hierarchical classcoordinate matrix, and inert-class attribute hierarchy of leukocyte images was analysizedto optimize attribute features. The70dimensions of original leukocyte images wasreduced to7using this method. The optimized attributes were discretized, and optimizedformal concept and attribute partial-ordered structure were established for actual leukocyte images dataset. Basing on the optimized attribute partial-ordered structure, sixclassification rules for six kinds of leukocytes was extracted.Finally, three leukocyte image classifiers were designed based on mixed-attributefusion in varied-space, which were multi-colour space (MCS), multi-scale space (MSS)and multi-level attribute space (MLS) respectively. Cross validation comparison of thethree classifiers and three traditional methods was carried out on actual dataset, and thecomparison results demonstrated that the classification accuracy followed the sequence ofMLS> MSS> MCS≈three traditional methods. Classification rules extracted fromattribute partial-ordered structure were well exploited in MLS classifier, anddecision-making examine using associated rules, error feedback and enhancedreprocessing were employed besides traditional classifiers. The experimental resultsindicated that the classification performance of leukocyte image classification methodbased on mixed attribute fusion in varied-space was remarkable, moreover, mixed attributefusion was capable of improving the classification accuracy. The MLS classifier also canbe applied for computer automated image identification in other fields.