DA-Capnet: Dual Attention Deep Learning Based on U-Net for Nailfold Capillary Segmentation

Abstract

Automatic nailfold capillary segmentation is a challenging task owing to noise and large variabilities in images caused by insufficient focusing and low visibility of the capillaries. This task can be useful to detect and estimate the severity of autoimmune diseases of connective tissues or learning the status of white blood cells based on the cells' blood flow on the nailfold capillary. Previous studies have addressed this task using manual, semi-automated, and automated segmentation method. However, further improvement is still required. With the recent progress of deep learning on medical imaging, we herein propose dual attention deep learning based on U-Net for nailfold capillary segmentation, named DA-CapNet. Our DA-CapNet improves the U-Net architecture by integrating a dual attention module that can capture a better representation of feature maps from input images. Furthermore, DA-CapNet is compared with three baselines: adaptive Gaussian algorithm, SegNet, the original U-Net. We experimentally demonstrate that our proposed method outperforms these baselines.