A new artificial-intelligence system has achieved accuracies of up to 92 percent when identifying musculoskeletal abnormalities in X-ray images, according to a study designed to address one of medical imaging’s most persistent problems: how to combine competing machine-learning models without forcing them to agree in every case. Rather than averaging the predictions of two deep-learning systems with fixed rules, the approach allows the models to adjust their influence from image to image. The researchers say this “dynamic ensemble” produced better results in most anatomical regions examined and could eventually help clinicians screen radiographs more rapidly, particularly in settings where radiologists are in short supply.

Musculoskeletal disorders encompass injuries and diseases affecting bones, joints, muscles, ligaments and related tissues. Their consequences can range from pain and temporary loss of movement to long-term disability, and X-ray imaging is often the first diagnostic tool used to detect fractures or other abnormalities. Yet interpreting these images is not always straightforward. Radiographs may contain subtle changes, unusual positioning, overlapping structures or image-quality problems that complicate diagnosis. The workload is also enormous, while trained radiologists are unevenly distributed around the world. Artificial intelligence has therefore attracted intense interest as a possible second reader—an automated system that flags suspicious images for closer human review rather than replacing clinical judgment.

The study, published in Medical & Biological Engineering & Computing, combines two different types of vision model: ConvNeXt Base and the Swin Transformer. ConvNeXt is a modern convolutional neural network, a class of model that analyzes images through layers of learned filters. Early layers detect simple patterns such as edges and textures; deeper layers assemble those signals into increasingly complex representations. Swin Transformer, by contrast, adapts the transformer architecture originally developed for processing sequences. It divides an image into local windows and calculates relationships among image patches, shifting the windows between layers so information can eventually flow across the entire radiograph. These different architectures may recognize different visual cues, creating an opportunity for one model to compensate when the other struggles.

Most medical-image ensembles combine models through a fixed average or a majority vote. If two models produce probabilities of abnormality, a conventional system might simply average those values, assigning each model equal influence regardless of the image. That assumption is convenient, but it may be wrong. A model could be more reliable for one anatomical region, image orientation or pattern of abnormality, while another could perform better under different conditions. The researchers instead used a Probability-based Dynamic Ensemble, or PDE, framework. It learns how much weight to assign to each model’s probability and also learns the decision threshold that separates an image classified as normal from one classified as abnormal. In this way, the final decision is not governed by a universal rule selected in advance.

The distinction between a probability and a final classification is important. A vision model may estimate, for example, that an image has a 0.60 probability of containing an abnormality. Whether that becomes a positive diagnosis depends on the threshold: a system using 0.50 would flag the image, while one using 0.70 would not. The optimal threshold depends on the balance between sensitivity and specificity. Sensitivity measures how effectively a system detects genuinely abnormal images, whereas specificity measures how successfully it avoids false alarms among normal images. Lowering the threshold may catch more abnormalities but can also increase unnecessary alerts. The PDE method attempts to optimize this trade-off using training data, while simultaneously learning how strongly each backbone model should contribute.

The researchers evaluated the system with the MURA dataset, a large collection of musculoskeletal radiographs assembled for abnormality detection. Their experiments covered seven anatomical regions and were conducted in two ways. In regional experiments, images were considered according to the body area represented, allowing the investigators to examine how performance varied across parts of the skeleton. In image-based experiments, the focus was on the individual radiograph as the unit of prediction. Before training, the X-rays were preprocessed and enhanced to make relevant visual information more consistent. The two backbone models were trained separately, and their dynamically generated probabilities were then used to construct and train the ensemble. This design enabled the combined system to exploit complementary recognition abilities rather than treating the models as interchangeable vote-casting machines.

Across the regional experiments, the dynamic ensemble recorded accuracies ranging from 82.6 to 92.0 percent, outperforming the individual models in most of the regions tested. Its image-based accuracy was 84.3 percent. Accuracy alone, however, does not reveal whether a model is missing dangerous abnormalities or generating too many false positives, so the study also included comparative statistical analyses of the results. The authors report that the learned probability fusion improved the balance between sensitivity and specificity, which is particularly important in clinical screening. A system that is highly sensitive but poorly specific could overwhelm radiologists with alerts; one that is highly specific but insensitive could allow genuine abnormalities to pass unnoticed. The reported gains suggest that dynamic weighting may offer a practical advantage over static ensemble strategies, although the study does not establish that the system is ready for independent clinical diagnosis.

The researchers also applied explainable-artificial-intelligence tools to examine the visual evidence influencing the model’s decisions. Such tools are intended to indicate which regions of an image contributed most strongly to a prediction, helping investigators identify whether the system is focusing on anatomically meaningful structures or on irrelevant features and imaging artefacts. Explainability cannot prove that an algorithm is correct, but it can expose suspicious behavior, support error analysis and give clinicians more context when reviewing an automated alert. This is crucial for medical deployment, where a highly accurate model may still be unsafe if its performance depends on scanner-specific markings, hospital workflows or other shortcuts that are absent from new clinical environments.

The findings add to a rapidly expanding effort to use deep learning in radiology, but they also underline why performance in a curated dataset is only an intermediate milestone. MURA is publicly available and provides an important benchmark, yet models trained on one dataset can lose accuracy when confronted with images from different hospitals, equipment, patient populations or acquisition protocols. The full study preview does not report prospective testing in routine clinical practice, nor does it demonstrate improved patient outcomes or reduced reporting times. The proposed system is therefore best understood as a decision-support framework: an algorithmic assistant that could prioritize examinations and help specialists inspect difficult cases. Before such a tool could influence care, it would need external validation, careful calibration, assessment across demographic groups, monitoring for changes in imaging practice and evaluation alongside radiologists. Even with those safeguards, the central idea is strikingly simple: instead of asking one artificial intelligence to see everything, let several models see the same X-ray—and allow the system to learn which set of eyes deserves greater trust each time.