Red Blood Cell Detection by the Improved Two-Layer Watershed Segmentation Method with a Full Convolutional Neural Network

Identifying cells in an image (cell segmentation) is essential for quantitative single-cell biology. The development of automatic medical image analysis of erythrocyte morphology has simplified the diagnosis of many diseases and combined a good reliability with high performance. However, in computational molecular biology, the elaboration of reliable and fully automated image analysis techniques for red blood cells imaging is still quite problematic. In this respect, quite lucrative is the watershed transformation concept, which can be utilized in image segmentation to generate partitions of the image corresponding to objects of interest. This approach is used in the current study, which presents a two-layer red blood cells detection framework, including a full convolutional neural network to extract the candidate cell regions with about 97% accuracy rate, and a novel labeled watershed method based on the morphology label and conditional skeleton extraction for improving the overlapped cells' segmentation. It is shown that the proposed method successfully extracts the red blood candidate regions, including single and overlapped ones. A comparative analysis of the proposed detection method and several other general methods was performed. The experimental results obtained indicate that the proposed framework is robust and accurate, with the cell extracting ratio over 87.76% accuracy.