This research aimed to study the application of deep learning to the diagnosis of rheumatoid arthritis (RA). Definite criteria or direct markers for diagnosing RA are lacking. Rheumatologists diagnose RA according to an integrated assessment based on scientific evidence and clinical experience. Our novel idea was to convert various clinical information from patients into simple two-dimensional images and then use them to fine-tune a convolutional neural network (CNN) to classify RA or nonRA. We semi-quantitatively converted each type of clinical information to four coloured square images and arranged them as one image for each patient. One rheumatologist modified each patient’s clinical information to increase learning data. In total, 1037 images (252 RA, 785 nonRA) were used to fine-tune a pretrained CNN with transfer learning. For clinical data (10 RA, 40 nonRA), which were independent of the learning data and were used as testing data, we compared the classification ability of the fine-tuned CNN with that of three expert rheumatologists. Our simple system could potentially support RA diagnosis and therefore might be useful for screening RA in both specialised hospitals and general clinics. This study paves the way to enabling deep learning in the diagnosis of RA.

这项研究旨在研究深度学习在类风湿关节炎（RA）诊断中的应用。缺乏用于诊断RA的明确标准或直接标记。风湿病学家根据基于科学证据和临床经验的综合评估来诊断RA。我们的新想法是将来自患者的各种临床信息转换为简单的二维图像，然后使用它们来微调卷积神经网络（CNN）以对RA或nonRA进行分类。我们将每种类型的临床信息半定量地转换为四个彩色正方形图像，并为每个患者将它们布置为一个图像。一位风湿病学家修改了每位患者的临床信息，以增加学习数据。总共使用1037张图像（252 RA，785 nonRA）对带有转移学习的预训练CNN进行了微调。对于独立于学习数据并用作测试数据的临床数据（10 RA，40 nonRA），我们将微调的CNN与三位风湿病专家的分类能力进行了比较。我们简单的系统可能支持RA诊断，因此对于在专科医院和普通诊所筛查RA可能有用。这项研究为在RA诊断中进行深度学习铺平了道路。