Artificial intelligence for nailfold capillaroscopy analyses - a proof of concept application in juvenile dermatomyositis.

Abstract

Biomarkers for idiopathic inflammatory myopathies are difficult to identify and may involve expensive laboratory tests. We assess the potential for artificial intelligence (AI) to differentiate children with juvenile dermatomyositis (JDM) from healthy controls using nailfold capillaroscopy (NFC) images. We also assessed the potential of NFC images to reflect the range of disease activity with JDM. A total of 1,120 NFC images from 111 children with active JDM, diagnosed between 1990 and 2020, and 321 NFC images from 31 healthy controls were retrieved from the CureJM JDM Registry. We built a lightweight and explainable deep neural network model called NFC-Net. Images were downscaled by interpolation techniques to reduce the computational cost. NFC-Net achieved high performance in differentiating patients with JDM from controls, with an area under the ROC curve (AUROC) of 0.93 (0.84, 0.99) and accuracy of 0.91 (0.82, 0.92). With sensitivity (0.85) and specificity (0.90) resulted in model precision of 0.95. The AUROC and accuracy for predicting clinical disease activity from inactivity were 0.75 (0.61, 0.81) and 0.74 (0.65, 0.79). The good performance of the NFC-Net demonstrates that NFC images are sufficient for detecting often unrecognized JDM disease activity, providing a reliable indicator of disease status. Proposed NFC-Net can accurately predict children with JDM from healthy controls using nailfold capillaroscopy (NFC) images. Additionally, it predicts the scores to JDM disease activity versus no activity. Equipped with gradients, NFC-Net is explainable and gives visual information beside the reported accuracies. NFC-Net is computationally efficient since it is applied to substantially downscaled NFC images. Furthermore, the model can be wrapped within an edge-based device like a mobile application that is accessible to both clinicians and patients.