Segmentation of White Blood Cells Using Deep Learning

Abstract: The white blood cell count and differential is an important part of diagnosing a number of medical conditions. Instead of doing this by manual microscopy, CellaVision’s technology has automated the process of finding and classifying white blood cells. To support a diagnosis it is desired that the system can produce cytoplasm-to-nucleus-ratio. This ratio is calculated from a segmented image where the pixels are labelled as background, cytoplasm, or nucleus. The system used today, using active contours, does not always produce perfect segmentations for all cells, and it would therefore be beneficial to improve the segmentation. Using machine learning, we have constructed a network for segmenting white blood cell images. This model, with some small modifications, produces both binary (cell and background) and multi-class (cytoplasm, nucleus and background) segmentations. The model is a U-net inspired by work previously done on other similar segmentation tasks. The network reached an IoU of 93.9% in the binary case, and in the muli-class case 82.8% and 94.5% for the cytoplasm and nucleus respectively. The main challenges were to separate neighbouring cells and cells in a cluster. Over all the network performed better than the active contour method in difficult images, and in cases where neither were good, the network was usually better. If the network was trained more on images that are difficult to segment, the resulting segmentations of these images could be improved.