Confidence scoring: Introducing the use case
Radiograph interpretation can be challenging: While some fractures are obvious at first glance, others are subtle and harder to detect. Their accurate reading might require additional clinical context, such as patient symptoms, the injury mechanism, or follow-up imaging. This variability in fracture appearance often leads to differences in interpretation, reflecting the complexity and nuance of musculoskeletal imaging.
To support clinicians in navigating these challenging cases, the RBfracture v2.5 introduces confidence scoring for each detected finding. This new feature goes beyond simply flagging potential fractures: It adds transparency by showing how certain RBfracture is in its assessment.
Displaying AI confidence empowers clinicians to calibrate their reliance on each finding, thereby enhancing RBfracture’s value as a decision-support tool in complex or uncertain scenarios.
RBfracture is intended to support the reading clinicians, and not replace them. All radiographs must still be reviewed by a clinician as part of the clinical process.
Confidence scoring in RBfracture's outputs
Interpreting bounding boxes and confidence scores
RBfracture offers two ways of visualizing its confidence: 1. Solid and dashed bounding boxes and 2. Percent-based confidence scores.
Depending on how you would like to use RBfracture to support your reading, you can configure RBfracture to show either one or both at once.
1. Solid and dashed bounding boxes
RBfracture is confident in the detected fracture, as indicated by the solid box. It is less confident about the dislocation. The lower confidence is indicated by a dashed box.
RBfracture categorizes trauma detections into two confidence levels:
- High Confidence: The vast majority of findings by RBfracture are displayed with a solid bounding box1. This indicates that RBfracture is highly confident that a finding is present. While these often correspond to clear findings, you still have to review all images and form your own decision.
- Low Confidence: Some findings are displayed with a dashed bounding box (- – -), which indicates that the finding has a moderate likelihood of being correct. RBfracture detected features suggestive of a fracture, but has a lower certainty.
1. Percent-based confidence scoring
RBfracture is 99% confident in the presence of a radius fracture, but only 93% confident about the ulna fracture.
For more detailed insight into RBfracture’s trauma detection, RBfracture can be configured to display a percent-based confidence score alongside each detected fracture.
This numerical value reflects how certain RBfracture is that a fracture is present at a specific location, based solely on imaging features.
- Higher scores (95–99%) indicate that the AI has seen strong radiographic evidence consistent with a finding2.
- Moderate scores (75–94%) suggest more subtle findings that the AI has flagged as potentially abnormal, but with less certainty.
These scores provide an additional layer of transparency, helping you decide how much weight to place on a given AI detection, and when to feel confident to overrule the AI.
Keep in mind, that the report is generated by AI and for decision support only. Always review the original images before making the final diagnosis.
Conclusion
In a time-pressured setting like the ED, every tool that improves diagnostic certainty is valuable. RBfracture not only highlights potential fractures and other trauma-related findings, it also tells you how confident it is in its assessment. That transparency empowers you to interpret AI predictions in the same nuanced way you already interpret radiographs: integrating imaging findings, clinical context, and professional judgment.
References
1 According to post-market surveillance, ~75% of all generated predictions are high-confidence. Information available on request.
2 According to internal test data, the high-confidence class has a PPV of 95%. Information available on request.
