Radiograph-Based Deep Learning Model to Support Finger Joint Selection for Ultrasound Examination in Rheumatoid Arthritis

Background/Objectives: Ultrasound is the standard imaging modality to evaluate the inflammatory changes in hand joints of rheumatoid arthritis (RA) patients. However, it is operator-dependent and takes a long time to examine. In this study, we developed a radiograph-based deep learning (DL) model to support prioritization of finger joints for ultrasound (US) examination in RA patients. Methods: In this retrospective study, hand radiographs from RA patients who underwent same-day US examination of bilateral finger joints were analyzed. A DL model was developed using hand radiographs from 270 patients (2043 finger joints) to estimate joint-level likelihood of inflammatory activity. US findings served as the reference standard for model training, while clinical findings of joint tenderness and swelling were incorporated as additional tabular inputs. Model performance was evaluated in a temporal-split test cohort consisting of 40 patients (270 joints) and compared with the performance of a clinical-only logistic regression model based on joint tenderness and swelling. Results: In the test set, the DL model demonstrated higher sensitivity (82.1% vs. 38.5%), negative predictive value (96.8% vs. 90.3%), and F1-score (69.6% vs. 48.4%) than the clinical-only model. Although the area under the receiver operating characteristic curve did not differ significantly between models (p = 0.43), precision–recall (PR) analysis showed superior performance of the DL model, with a higher area under the PR curve (0.625 vs. 0.540). At the threshold maximizing the F1-score, DL-assisted triage reduced the number of finger joints selected for US examination by approximately 80%. Conclusions: A radiograph-based DL model can support efficient prioritization of finger joints for US examination in RA, offering a practical approach to enhance joint-level US triage in routine clinical practice.