Segmentation of white blood cells and comparison of cell morphology by linear and naïve Bayes classifiers

• DOI: 10.1186/s12938-015-0037-1
• Publication Date: 2015-06-29T09:09:15Z
• Authors: Jaroonrut Prinyakupt; Charnchai Pluempi...
• Keywords: Digital Pathology; Feature (linguistics); White blood cell

Blood smear microscopic images are routinely investigated by haematologists to diagnose most blood diseases. However, the task is quite tedious and time consuming. An automatic detection classification of white cells within such can accelerate process tremendously. In this paper we propose a system locate images, segment them into nucleus cytoplasm regions, extract suitable features finally, classify five types: basophil, eosinophil, neutrophil, lymphocyte monocyte. Two sets were used in study's experiments. Dataset 1, collected from Rangsit University, normal peripheral slides under light microscope with 100× magnification; 555 601 captured Nikon DS-Fi2 high-definition color camera saved JPG format size 960 × 1,280 pixels at 15 per 1 μm resolution. dataset 2, 477 cropped cell downloaded CellaVision.com. They 360 363 pixels. The resolution estimated be 10 μm. proposed comprises pre-processing step, segmentation, feature extraction, selection classification. main concept segmentation algorithm employed uses cell's morphological properties calibrated real relative image combined thresholding, operation ellipse curve fitting. Consequently, several extracted segmented regions. Prominent then chosen greedy search called sequential forward selection. Finally, set selected prominent features, both linear naïve Bayes classifiers applied for performance comparison. This was tested on slide two datasets. comparison performed: automatically results compared ones obtained manually haematologist. It found that method consistent coherent datasets, dice similarity 98.9 91.6% average respectively. Furthermore, overall correction rate phase about 98 94% models, system, based morphology its characteristics, different datasets fast, robust, efficient coherent. Meanwhile, types shows high sensitivity slightly better classifier.