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

Area of Science:

• Artificial Intelligence in Medical Imaging
• Rheumatology
• Diagnostic Imaging

Background:

• Ultrasound (US) is standard for evaluating rheumatoid arthritis (RA) joint inflammation.
• Current US examination is operator-dependent and time-consuming.
• Need for efficient prioritization of joints for US in RA patients.

Purpose of the Study:

• Develop and evaluate a radiograph-based deep learning (DL) model.
• Support prioritization of finger joints for US examination in RA.
• Enhance joint-level US triage in clinical practice.

Main Methods:

• Retrospective analysis of hand radiographs from RA patients with same-day US.
• Developed a DL model using radiographs and clinical data (tenderness, swelling).
• Compared DL model performance against a clinical-only logistic regression model.

Main Results:

• DL model showed higher sensitivity (82.1% vs. 38.5%) and F1-score (69.6% vs. 48.4%) than the clinical model.
• DL model demonstrated superior area under the precision-recall curve (0.625 vs. 0.540).
• DL-assisted triage reduced US examination of finger joints by approximately 80%.

Conclusions:

• A radiograph-based DL model effectively prioritizes finger joints for US in RA.
• This AI approach offers a practical method to improve US triage efficiency.
• The model supports routine clinical practice for RA joint assessment.