Segmentation of Overlapping Red Blood Cells for Malaria Blood Smear Images by U-Net Architecture

Abstract

Automated segmentation of red blood cells is a widely applied task in order to evaluate red blood cells for certain diseases. Counting of malaria parasites requires individual red blood cell segmentation in order to evaluate the severity of infection. For such an evaluation, correct segmentation of red blood cells is required. However, it is a difficult task due to the presence of overlapping red blood cells. Existing methodologies employ preprocessing steps in order to segment red blood cells. We propose a deep learning approach that has a U-Net architecture to provide fully automated segmentation of red blood cells without any initial preprocessing. While red blood cells were segmented, irrelevant objects such as white blood cells, platelets and artifacts were removed. The network was trained and tested on 5600 and 600 samples respectively. Segmentation of overlapping red blood cells was achieved with 93.8% Jaccard similarity index. To the best of our knowledge, our results surpassed previous outcomes.