Title: Mimicking Clinical Workflow: Anatomical Spatial Priors for Accurate Cephalometric Landmark Localization

Abstract: While medical professionals rely on a structured anatomical process when tracing cephalometric radiographs, existing computational systems have failed to incorporate this methodology. To address this gap, we introduce a five-stage pipeline that generates confidence-weighted spatial priors based on anatomical structure. These priors guide the training of an HRNet-W32 model, resulting in a mean radial error of 1.04 mm across 25 landmarks evaluated on 1,502 radiographs captured by more than seven different imaging devices.

Our analysis utilizes a training-inference prior matrix to isolate the underlying mechanism. The data reveals that anatomical priors limit the discrepancy between validation and test performance to just 1%, compared to an 88% gap observed in models without priors (which yielded a 1.94 mm error), even though both models achieved identical convergence during validation. This matrix demonstrates several key findings: all trained models operate independently of inference conditions; simply expanding the architecture offers no performance gain; random priors provide only partial and unstable improvements (averaging 1.72 mm); and only image-specific, anatomically accurate priors achieve the 1.04 mm accuracy. These effective priors act as a regularizer during training, eliminating the need for automated prior generation at the deployment stage.

We present converging evidence through five-fold cross-validation (p=0.0015) and patient-level permutation testing, which showed no significant reversals (ICC > 0.95). Furthermore, cross-domain experiments involving echocardiography, cervical spine, and hand radiography support the hypothesis that the efficacy of these priors is directly related to the spatial entropy of the landmark distribution.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC