NFCMTL: Auto NailFold Capillaroscopy through a Multi-Task Learning Model

Yingke Ding, Jiankai Tang, Wanying Mo, Tianruo Rose Xu, Yuanchun Shi, Yuntao Wang

Keywords: Nailfold Capillaroscopy, Multitask Learning, Vision Transformer

TL;DR: Novel multitask learning method for nailfold capillaroscopy images

Abstract: Nailfold capillaroscopy is a non-invasive technique for assessing microvascular health by visualizing capillaries in the nailfold, playing a key role in diagnosing vascular and autoimmune diseases. We propose a novel machine learning approach for nailfold analysis, introducing an advanced multi-task learning model that jointly performs capillary segmentation, classification, and keypoint detection within a unified architecture. Using a large public dataset with reorganized keypoint annotations, our approach improves precision and efficiency in feature detection while simplifying the conventional multi-stage pipeline. By leveraging multi-task optimization, the model achieves state-of-the-art performance comparable to existing methods. This work advances nailfold imaging by providing an accurate, streamlined solution for automated, non-invasive microvascular diagnostics. Code is available at https://github.com/thuhci/NFCMTL.

Submission Number: 16