Background/Purpose: Nailfold capillaroscopy is a key tool in the early diagnosis of systemic sclerosis and related disorders. However, incorrect image acquisition can lead to misclassification of capillaroscopic patterns, potentially delaying diagnosis and affecting clinical decisions.To support clinical reliability we developed a lightweight, real-time system for automatic quality assessment of nailfold capillaroscopy (NFC) images. The system supports exclusion of invalid images from analysis and reporting, provides immediate feedback during image acquisition, and aids clinician training by detecting typical image acquisition problems during a capillaroscopy examination.

Methods: A small and efficient deep learning model based on the ConvNeXt V2 architecture was trained to identify frequent NFC image issues: low magnification, blurring, reflections, incorrect placement, underexposure, and invalid content. For valid images (those without problems), the model also estimates diagnostic usability as low quality or acceptable quality.Low quality images were defined as diagnostically acceptable but with imperfect focus or suboptimal capillary visibility, while acceptable quality images had clear visualization suitable for confident pattern recognition.The model was trained on a dataset of 10,866 annotated images. An image may have multiple problem annotations, but quality grading (low or acceptable) was applied only to images without any problems. The number of annotations per annotation type are shown in Table 1.To ensure clinical reliability, model evaluation metrics were computed at a confidence threshold of 0.85.

Results: The model achieved exceptional performance in detecting low magnification images, with an F1-score of 0.99, and invalid content, with an F1-score of 0.96, two of the most critical categories for ensuring diagnostic image quality. It also showed strong performance in identifying blurry images (F1 = 0.84) and those with excessive reflections (F1 = 0.83).In terms of quality grading of valid images, the model effectively distinguished between low and acceptable quality among valid images, achieving an F1-score of 0.8316 for low-quality cases. The rest of performance metrics are shown in Table 2.

Conclusion: This real-time model based on ConvNeXt V2 helps ensure that nailfold capillaroscopy examinations are useful and valid. It automatically detects common problems in images and checks whether valid images have enough quality for diagnosis. Because it is fast and lightweight, it can be used during image capture to give immediate feedback, avoid invalid images, and improve the consistency of capillaroscopies. This makes the process more efficient and helps both clinicians and researchers achieve better, more reliable results as well as being a useful tool in the training of new physicians.

Table 1. Number of annotations per class

Table 2. Performance metrics per class

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/automated-feedback-and-quality-control-in-nailfold-capillaroscopy-a-tool-for-clinical-and-educational-use/