Unsupervised segmentation of leukocytes images using thresholding neighborhood valley-emphasis

Abstract

Blood smear image analysis is essential to correlate the amount of leukocytes in these images with malignancies such as the leukemias. Techniques of digital image processing can be used to aid pathologists in this analysis, leading to appropriate treatments for the patient. This paper presents an unsupervised segmentation method for the nuclear structures in leukocytes. Deconvolution was used to split the Giemsa stain components and the regions of interest were selected using a thresholding algorithm called Neighborhood Valley-emphasis. A postprocessing approach based on morphological operators was applied in these detected structures. The proposed algorithm was tested on 367 images containing leukocytes and other blood structures. A performance analysis was conducted through the Jaccard and accuracy metrics featuring results of 89.89% and 99.57%, respectively. Such results were compared to other published articles and this was considered the most promising method.