Classification of Overlapping Red Blood Cells Using Image Segmentation and Convolutional Neural Network

Abstract

Machine vision is an analytical technique widely used to identify blood cells, considering both qualitative and quantitative attributes. Qualitative studies include differentiating red blood cells and white blood cells. Quantitative studies include counting the number of cells in a particular image. However, the problem is that the phenomenon of overlapping, especially in red blood cells images, often occurs, which causes the accuracy of image processing to be disrupted. In the present study, the image segmentation process for image red blood cells using convolutional neural network algorithms has been evaluated to predict single and overlapping classification groups on red blood cells. A total of 100 public red blood cells images were used in this study by treating 5 data splits (calibration and testing), including 80:20, 75:25, 70:30, 65:35, and 60:40. Each cell in the red blood cells image will be classified into a single red blood cells group and an overlapping red blood cells group. Probability statistical tests and ANOVA were used to obtain the best preprocessing strategy for each classification parameter, including precision, recall, F1-score, and accuracy. The results generally show that image segmentation can classify single and overlapping red blood cells with acceptable accuracy.