Cervical Vertebrae Context Improves Bone-Age Estimation Across Sites
A multicenter deep learning study suggests that bone-age assessment from lateral cephalograms becomes more resilient when the model is allowed to see more than the vertebrae alone. In this retrospective dataset, investigators compared three cervical vertebrae-centered input strategies for estimating bone age from lateral cephalograms and asked whether nearby upper-airway anatomy could add useful signal. The central issue was cross-site generalizability, because models that perform well in one institution can lose accuracy when scanner settings, patient positioning, or image characteristics shift at another center.
The study analyzed 1,100 paired lateral cephalograms and hand–wrist radiographs from two institutions in children and adolescents aged 4 to 18 years. Reference bone ages came from AI-derived surrogate labels generated with commercial hand–wrist bone-age software. Second through fourth cervical vertebrae and the upper airway were manually outlined to create three strategies: CV masking, combined CV and airway masking, and CV-centered cropping without masking. Those design choices matter because cephalometric models often depend on anatomy that is visually adjacent to the target region, and strict masking can strip away contextual cues that a network may have been using to stabilize its estimate.
The result was not a simple story of one approach dominating every analysis.
Under internal evaluation, performance looked broadly similar across the three strategies. Once the model faced an external test set, the picture changed. Overall errors increased substantially across the board, but the deterioration was worst with CV-only masking, which produced a mean absolute error of 1.91 years. Airway-inclusive masking reduced the external mean absolute error to 1.72 years, and unmasked CV-centered cropping performed best at 1.67 years. The relative ranking points to a familiar problem in medical imaging AI: when a model is trained on tightly constrained anatomy, it may become brittle when the visual distribution changes outside the training institution.
Why Surrounding Anatomy Appears to Help the Network
The likely explanation is straightforward from a machine-learning perspective and plausible from a radiologic one. Cervical vertebral maturation is the intended target, but lateral cephalograms contain neighboring anatomy that may encode positioning, contour continuity, and developmental context. By preserving the upper airway or leaving the vertebral region uncropped, the model retains spatial relationships that can improve stability when the image comes from a different scanner, a different acquisition protocol, or a different patient mix. In effect, the network is not limited to isolated vertebral pixels; it can also use the local anatomical frame around them.
Interpretability analysis supported that reading. The attention maps mainly emphasized cervical vertebral structures, which suggests the model did not abandon the intended biological target. At the same time, Grad-RAM for airway-inclusive inputs showed prominent activation in the airway region. That pattern indicates the surrounding airway was not simply ignored as irrelevant background. Instead, it likely contributed as a contextual cue when the model inferred skeletal maturation from the cephalogram.
The external test set exposed the real constraint in this kind of work. Cross-site performance eroded even though the internal evaluation was comparatively stable, a sign that model behavior can look reliable until it is tested under different image conditions. For clinical deployment, that is the relevant bar. A bone-age tool that depends too rigidly on a narrow visual crop may work well in development and then drift when used elsewhere.
Bone-Age AI Still Depends on the Quality of Its Reference Labels
Another practical issue sits beneath the modeling itself: the study used AI-derived surrogate labels from commercial hand–wrist bone-age software rather than direct manual skeletal age adjudication. That choice allows large-scale analysis, but it also means the regression target is itself an algorithmic estimate. In pediatric dentistry and orthodontic planning, bone age is often used to support growth assessment and treatment timing, so the calibration of any automated method matters. If the reference label carries error, the model can learn a polished version of the wrong answer.
Even so, the comparison across input strategies remains informative. All three methods were tested within the same multicenter cohort, and the consistent finding was that preserving anatomical context improved robustness when the model moved outside its familiar environment. Strict vertebra-only masking produced the weakest external result, which argues against over-pruning images in an effort to remove every non-target structure. For cephalometric AI, useful context may be a feature rather than a flaw.
What the Multicenter Design Adds to Orthodontic Imaging AI
The multicenter design gives this work more weight than a single-institution proof of concept. Pediatric dental imaging data are particularly vulnerable to distribution shift because patient age, growth stage, imaging technique, and annotation standards can vary substantially from one center to another. By testing internal and external performance separately, the study exposed the difference between in-house accuracy and real-world portability. In that sense, the external test set was the most clinically meaningful part of the analysis.
A few implications follow directly. First, cephalogram-based bone-age systems may benefit from C V-centered cropping that preserves surrounding anatomy instead of aggressive masking. Second, airway-inclusive processing may provide additional context without losing focus on the vertebrae. Third, external validation should remain mandatory before any such tool is considered for routine use, because the most appealing internal metrics may not survive site-to-site variation.
The study does not establish a universal best preprocessing method for all cephalometric AI tasks, and it does not show whether the same advantage would appear in other age ranges, other populations, or other imaging platforms. It does show something narrower and more useful: in pediatric bone-age estimation, models that can see the cervical vertebrae in context were more robust than models forced to look at isolated vertebrae alone. For a workflow built around growth assessment, that extra resilience could determine whether an algorithm remains a research demonstration or becomes a dependable clinical aid.