Observation and density estimation of a large number of skin capillaries using wide-field portable video capillaroscopy and semantic segmentation.

Significance: Skin capillaries are non-invasively observable; their structure and blood flow can reflect tissue and systemic conditions. Quantitative analysis of video-capillaroscopy images yields novel diagnostic methods. Because the capillary structure is heterogeneous, analyzing more capillaries can increase the evaluation reliability. Aim: We developed a system that can observe and quantify numerous capillaries and verified the performance on human skin. Approach: We developed a portable video-capillaroscope with a spatial resolution higher than 3.5 µm and a wide field of view (7.4 mm×5.5 mm) and a method to evaluate capillary numbers and areas using U-Net. The model was trained and tested with 22 and 11 cropped images (2.4 mm×1.9 mm) obtained from 11 participants, respectively. They were then applied to the 7.2 mm×5.3 mm images from four participants. Segmentation results were compared to ground-truth at the pixel level and capillary-region level. Results: Over 1000 capillaries were simultaneously observed using the proposed system. Although pixel-level segmentation performance was low [intersection over union (IoU) = 24.5%], the number and area could be estimated. These values differed among four participants and seven sites, and they changed after skin barrier destruction. Conclusions: The proposed system allows for observing and quantifying numerous skin capillaries simultaneously, suggesting its potential for evaluating tissue and systemic conditions.