Background/Purpose: Rheumatology musculoskeletal ultrasound (MSKUS) is a powerful technique to identify and differentiate arthritidies joint pathology, yet MSKUS is highly dependent on user experience. Techniques improve ultrasound training, imaging explainability, and aid clinical interpretation can increase the utility of MSKUS. We have begun work on SynovAI, an innovative, next-generation artificial intelligence (AI) framework designed to augment MSKUS interpretation and diagnostic accuracy. The goal SynovAI is to develop a framework which can identify, quantify, and highlight relevant pathologic features on MSKUS and provide likely rheumatic diagnoses based on the imaging alone, or in combination with clinical data inputs.

Methods: SynovAI framework integrates a suite of pioneering deep learning (DL) strategies designed for maximal diagnostic efficiency and interpretability. The framework begins with (1) an attention-enhanced U-Net architecture that adeptly segments synovial regions of interest (ROI), significantly reducing diagnostic noise. Subsequently, (2) an innovative dual-branch neural network merges global semantic insights derived from a Convolutional Block Attention Module (CBAM)-enhanced ResNet architecture with subtle textural details extracted by a unique diffusion-inspired model, enabling precise discrimination between B-mode ultrasound features. Then, (3) an advanced, attention-based adaptive feature fusion mechanism synthesizes these complementary features into highly robust representations.Pilot evaluation was conducted on a retrospective clinical dataset of 588 B-mode ultrasound images from dorsal longitudinal metacarpophalangeal (MCP) views, all of which were interpreted by an expert ultrasonographer for the presence or absence of synovial hypertrophy. The dataset was split into 424 training images, 105 validation images, and 59 held-out test images, and reflected real-world clinical variability, including imaging artifacts and inconsistent probe angulation. Training included mixup augmentation, focal loss, gradient clipping, and cosine annealing with warm restarts. Model predictions were further refined via ensemble learning, adaptive thresholding, and test-time augmentation (TTA). Interpretability was enhanced with Grad-CAM visualizations.

Results: On the test set, SynovAI achieved 93.2% accuracy and a macro-averaged F1-score of 0.93, with F1-scores of 0.94 for normal and 0.92 for inflammatory joints. These results highlight SynovAI’s strong and consistent performance in detecting early synovial abnormalities in routine ultrasound imaging.

Conclusion: In this early proof of concept work, SynovAI could identify synovial changes with high accuracy by combining state-of-the-art deep learning methodologies with unparalleled interpretability. We plan to explore identification of associated bony, tendinous, and soft tissue features to synthesize in a predictive model to identify likely diseases, with the addition capability of handling secondary clinical inputs and providing explainable imaging outputs.

Performance of synovial AI on the test set of metacarpophalangeal joints with varying degrees of synovial hypertrophy (top). Image segmentation, grad-CAM heat map, and visualization of the region of interest identified by SynovAI (bottom)

views of the dorsal MCP with normal synovium (A, arrows) and abnormal synovial hypertrophy (B-D, arrowheads). mc= metacarpal head; et = extensor tendon; pp = proximal phalanx.

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/synovai-a-revolutionary-ai-framework-for-enhanced-detection-and-differentiation-of-rheumatoid-arthritis-osteoarthritis-and-spondyloarthritis-via-advanced-ultrasound-imaging/