Automated elbow ultrasound image recognition: a two-stage deep learning system via Swin Transformer.

Background: Ultrasound imaging is pivotal for point of care non-invasive diagnosis of musculoskeletal (MSK) injuries. Notably, MSK ultrasound demands a higher level of operator expertise compared to general ultrasound procedures, necessitating thorough checks on image quality and precise categorization of each image. This need for skilled assessment highlights the importance of developing supportive tools for quality control and categorization in clinical settings. To address these challenges, we aim to develop and evaluate an automated system designed to assess whether an ultrasound image meets established standards and to identify its specific category, thereby assisting clinicians in improving diagnostic efficiency and accuracy in MSK ultrasound.

Methods: We proposed a two-stage system built upon a Swin Transformer-based Unet model. Initially, an ultrasound image of the elbow is categorized, following which its quality is evaluated to determine if it meets the established standard criteria set by experienced radiologists.

Results: The proposed method surpassed other approaches in performance, with the Swin Transformer, serving as the backbone, outperforming conventional convolutional neural network (CNN) models. Additionally, when utilizing the same backbone, the two-stage model exceeded the capabilities of the single-stage model. Notably, the accuracy of our model exceeded 97%, with each category achieving over 95%. Additionally, processing time was significantly reduced from over 16 minutes to less than 15 seconds. Furthermore, this model reduces the workload for experienced radiologists, demonstrating its efficiency and effectiveness. This allows human experts to focus more on rare and challenging cases that machine learning struggles to solve, and to create high-quality labels for the models to learn from.

Conclusions: We have introduced an elbow ultrasound image recognition system leveraging the Swin Transformer, achieving competitive outcomes in both categorization and standard verification. This approach significantly reduces diagnostic time, allowing radiologists to focus more on scientific validation rather than image analysis.