Automated Shoulder Radiograph Quality Review to Support Efficient Workflow in the Emergency Department: The SQUIRE Deep Learning Ensemble.
Authors
Affiliations (7)
Affiliations (7)
- Department of Computer Science, University of California, Irvine, Irvine, CA, USA. [email protected].
- Department of Radiological Sciences, University of California Irvine School of Medicine, Irvine, CA, USA. [email protected].
- University of California Irvine School of Medicine, Irvine, CA, USA. [email protected].
- Department of Orthopedic Surgery, University of California Irvine School of Medicine, Irvine, CA, USA.
- Department of Computer Science, University of California, Irvine, Irvine, CA, USA.
- Department of Radiological Sciences, University of California Irvine School of Medicine, Irvine, CA, USA.
- University of California Irvine School of Medicine, Irvine, CA, USA.
Abstract
Suboptimal shoulder radiographs are a persistent challenge, leading to unnecessary radiation exposure, workflow delays, and increased costs from repeat imaging. Repeat imaging is most often required due to anatomy cutoff or suboptimal patient positioning, underscoring the need for targeted feedback and decision-support tools to improve image acquisition consistency. This study aimed to develop and validate an automated approach for assessing shoulder radiograph quality under realistic deployment conditions. We introduce new expert-verified quality annotations for 732 shoulder radiographs from the publicly available MURA dataset-representing, to our knowledge, the first publicly released labels specifically on shoulder radiograph acquisition quality. Labeling focused on Grashey (True AP) and axillary views, the most frequently repeated shoulder radiograph views in clinical practice. Each radiograph was categorized as diagnostically adequate or requiring repeat imaging. Using these labeled images, we developed SQUIRE (SHoulder QUality Image REviewer), a real-time heterogeneous ensemble model composed of lightweight (< 50M parameters) architectures selected through intra- and inter-architectural five-fold cross-validation. Internally, SQUIRE achieved an average F1 score of 0.943 and accuracy of 0.941, outperforming individual architectures and demonstrating reduced variability across folds. External validation on a temporally and institutionally distinct dataset revealed expected performance degradation due to domain and annotation shift. Nevertheless, SQUIRE maintained robust discrimination capacity in zero-shot settings (ROC AUC = 0.82) and few-shot adaptation (ROC AUC = 0.85). Comparison with a frozen medical foundation model further highlighted the advantage of domain-specialized, lightweight architectures for acquisition-quality assessment. By openly releasing these annotations and code, we establish a reproducible benchmark for low-latency, point-of-care radiographic quality assessment.