You are here
Detecting and Segmenting White Blood Cells in Microscopy Images of Thin Blood Smears [Poster].
Automatic analysis of microscopic images of blood smears with parasite-infected red blood cells can assist field workers in screening and monitoring malarial infection. Millions of blood slidesare manually screened for parasites every year, which is a time consuming and error prone process. We have developed a smartphone-based software to perform this task using machine learning and image analysis methods for counting infected red blood cells automatically. The software runs on a standard Android smartphone attached to a microscope by a lowcost adapter. The smartphone’s built-in camera views images of thin blood smear slides through the eyepiece of the microscope. Our method first detects and segments red blood cells in these images. However, the presence of white blood cells (WBCs) adversely affects the accuracy of red blood cell detection and segmentation since WBCs are often mistaken for red blood cells by current automatic cell detection methods. Therefore, a pre-processing step for WBC elimination is necessary, which requires WBC detection and segmentation methods. Segmentation of WBCs is a complex process due to the morphological diversity of WBCs and staining differences. We introduce an algorithm that successfully detects white blood cells in microscopic images of thin blood smears following a range filtering approach that accurately estimates the boundary of each cell using a level-set algorithm. Our method is capable of detecting different types of white blood cells with different shades of staining. We test our algorithm on more than 1300 thin blood smear images exhibiting about 1350 WBCs, which are manually annotated by a professional slide reader. For cell detection, we achieve 96.4% precision, 98.4% recall, and 97.4% F1-score. For cell segmentation, we assess the performance by a pixel-wise comparison of the manual segmentations with our machine segmentations, which has resulted in an overlap of 82% based on the Jaccard similarity index. These results are promising outcomes for automatic WBC analysis, which lead to higher accuracy in the counting of red blood cells and a concomitant improvement in screening performance.