Research On Recognition Of Bone Marrow Cells Based On Extreme Learning Machine

The bone marrow is the major hematopoietic organs in human.There is a great variety of cells in it. Many hematologic diseases can bediagnosed and differentiated by the classification of the bone marrowcells. Microscopic examination is the main diagnosis method, but themanual operation is hard, it mingles too many subjective factors. Socomputer analysis and recognition microscopic image is greatlysignificant. The machine learning algorithm of extreme learningmachine can automatic analysis and recognition the bone marrow cells,experiment proves the validity of this method. The main contents are asfollows:(1) The collected bone marrow cells image is segmented by theGrabCut algorithm based on the saliency of refined context aware(RCA-GrabCut) firstly. This algorithm can overcome the disadvantageof traditional GrabCut algorithm that requires manual interaction.GrabCut algorithm is suitable for segmenting the bone marrow cellsimage with complicated background. The proposed RCA algorithmobviously improves the computational speed, which is better than thetraditional context aware saliency algorithm. Compared with otheralgorithms, the overall error rate of the RCA-GrabCut algorithm is lowand its effect is good. Aiming at the conglutinate or overlapphenomenon existed among the cells in the cell image, the dilationoperator and erosion operator are used. When a cell is extracted fromthe bone marrow cells image, the cell image is threshold by otsu methodand the cell nucleus area is separated from the background.(2) The characteristics values of the morphology, optical density,texture and fractal extracted from the bone marrow cell image arecombined. Firstly the quantitive indices of cell perimeter, area andnuclear branch numbers are extracted from the single bone marrow cell segmentation, which is describing cell morphology. Then the opticaldensity characteristics based on statistical color characteristicsinformation are extracted. Finally the texture and fractal featureparameters of bone marrow cells are extracted. The number ofmorphology, optical density, texture and fractal is thirty-nine in total,and the data is stored in certain format.(3) An ensemble of extreme learning machine algorithm based onthe cellular automata (CA-E-ELM) is proposed for the classification ofbone marrow cells. Firstly, the principle and flow of ELM areintroduced. Then the experimental results show that this algorithm hasfast learning speed and good generalization performance withoutadjusting any parameters during run-time compared with BP neuralnetworks and support vector machines. For the instability ofperformance of single classifier, the classifiers ensemble method isadopted. Accurate and discrepant classifier is obtained by the cellularautomata method through disturbing the training set. The CA-E-ELMalgorithm is applied in bone marrow cells classification reached as highas97.43%. Moreover, it effectively solves the disadvantage ofinstability for the neural network classifier.