| Sumario: | Background: Rheumatoid arthritis (RA) is a chronic inflammatory disease that damages hand and wrist joints, leading to pain, disability, and reduced quality of life. Radiographic assessment plays a key role in diagnosis, but it is subjective and dependent on the clinician's experience. Deep learning–based systems offer the potential for faster, more objective, and more consistent evaluation. Objectives: This study aims to develop an attention-based deep learning model for the automated diagnosis of RA from hand and wrist radiographs and to demonstrate that high diagnostic performance can be achieved even with a limited data set. Design: Retrospective observational study evaluating an attention-based deep learning model for automated diagnosis of RA from hand and wrist radiographs. Methods: Radiographs from 311 RA patients and 259 healthy controls collected between September 2018 and September 2024 were analyzed. Individuals with other conditions causing hand deformities were excluded. The data set was divided into training (n = 325), validation (n = 142), and test (n = 50) sets. DenseNet121 and DenseNet169 architectures were combined with an attention mechanism to highlight RA-specific structural changes. Despite the relatively small data set, data augmentation and attention modeling were used to improve robustness. Results: The proposed model achieved 88% accuracy, 84% precision, and 91% recall, demonstrating strong diagnostic capability with limited training data. Initial clinical testing suggests that the model can support radiologists by providing consistent and objective assessments. Conclusion: This attention-based deep learning approach shows promise as an effective, reliable, and efficient tool for the automated diagnosis of RA. The ability to achieve high performance with limited data highlights its potential for real-world clinical adoption, particularly in resource-constrained environments.
|
|---|