Optimization of overlapping dendritic cell segmentation in phase contrast microscopy images

Abstract

The segmentation of Dendritic Cell (DC) from clumps of overlapping Peripheral Blood Mononuclear Cell (PBMC) in Phase Contrast Microscopy (PCM) images is notoriously challenging for an automated image analysis system. This problem is encountered due to the presence of shade off effect and halo region in the image. In order to improve the performance of DC classification, the methods in pre-processing are enhanced and analysed. The images undergo image normalization process to remove uneven illumination. Initially, Local Contrast Threshold (LCT) has been applied in preprocessing. However, it results in low performance of DC segmentation and identification. Therefore, a hybrid of low and high sigma in Gaussian kernel filtering with Local Adaptive Threshold (H-GLAT) through logical operator AND are proposed. Following that, halo removal is applied to eliminate halo region and post-processed by morphological operators to discriminate the cells from the background. The quantitative assessment demonstrates that proposed framework can successfully address these imaging artifact issues. The test results show that the H-GLAT method is better than the LCT that applied in previous work with classifier performance of 76%, 93.3% and 99.5% precision, recall and accuracy respectively.