Segmenting nailfold capillaries using an improved U-net network.

To assess the microcirculation in a patient's capillaries, clinicians often use the valuable and non-invasive diagnostic tool of nailfold capillaroscopy (NC). In particular, evaluating the images that result from NC is particularly important for diagnosing diseases in which the capillary morphology is altered. However, NC images are generally of poor quality, such that analyzing them is difficult and time consuming. Thus, the purpose of this work was to determine a way to segment the capillaries in poor-quality NC images accurately. To do this, we proposed using a deep neural network with a Res-Unet structure. The network combines the residual network (ResNet) and the U-Net to establish an encoding-decoding network and to deepen the layers in the network to preserve the features of the deep layer. The network was trained on 30 nailfold capillary images to discriminate the pixels belonging to capillaries, and it was then tested on a dataset consisting of 20 images to achieve a binarized map. The mean accuracy was 91.72% and the mean Dice score was 97.66% compared to the ground truth, which indicates that using Res-Unet to perform capillary segmentation in NC images had good performance.