Routine CBCT hides invisible soft-tissue disease: a frozen BertsWin encoder with physics-guided pretraining lifts MRI-verified joint effusion above chance
Synopsis
Cone-beam CT (CBCT) is built to resolve bone, so soft-tissue pathology usually needs a separate MRI. In 99 patients and 212 paired temporomandibular joints imaged with both CBCT and MRI, this study used physics-guided 3D self-supervised pretraining, a BertsWin encoder trained with the PhysLoss objective and the GCond optimizer, then froze the encoder and trained only a lightweight classifier. Detection of MRI-verified synovial effusion reached statistically robust discrimination, with the lower confidence bound still above chance after a conservative degrees-of-freedom adjustment, while the same architecture pretrained with a standard loss or unmasked variance decomposition fell to random guessing.
Interpretation
Routine CBCT scans physically retain soft-tissue pathology signals that are invisible to the eye, and those signals can be extracted for opportunistic screening instead of requiring a prior MRI referral. CBCT had been treated as a frontline modality for mineralized structures, with non-calcified findings such as joint effusion and synovial inflammation confirmed by MRI; inferring soft-tissue status from surrogate joint-space geometry often gave specificity below 50%. On a paired CBCT–MRI cohort of 99 patients and 212 temporomandibular joints, detection of MRI-verified synovial effusion reached statistically robust discrimination, and the lower 95% confidence bound of ROC AUC stayed clearly above chance under a conservative degrees-of-freedom adjustment.
A frozen 3D encoder keeps the soft-tissue signal under very few paired samples only when a physical prior is written into the pretraining objective. A standard L2 loss and an unmasked multi-component variance (MVC) objective let the encoder spend capacity on high-contrast cortical bone boundaries, where soft-tissue gradients are swamped. The same BertsWin architecture pretrained with the standard loss or MVC landed at chance-level AUC for effusion, with validation Youden indices indistinguishable from random guessing; only the PhysLoss plus GCond combination produced a statistically significant result.
The classifier receives only a small latent sub-volume around the condylar zenith rather than the whole peri-articular volume. Feeding the full encoder feature map to the classifier let the spatial variance of surrounding cranial and mandibular bone overwhelm the weak intra-articular signal, and preliminary experiments gave no statistically conclusive predictions. Automatic localization of the mandibular condylar zenith with an Attention U-Net head followed by cropping to a local token sub-grid discarded 81.3% of background osseous variance, and the zenith coordinates were shared across all compared encoders so the comparison isolated representation quality.
Perspective
The results come from a single-center retrospective cohort of 99 patients and 212 temporomandibular joints, with paired scans restricted to a 30-day interval and CBCT volumes spanning six scanner models; the crop covers only the peri-articular volume, excluding extra-articular soft tissues such as the lateral pterygoid muscle.
The authors state explicitly that the failure of standard-loss or MVC encoders should not be read as those representations being unable to encode soft-tissue features, since they might succeed with several thousand paired volumes; even the best configuration showed localized noise overfitting on some folds, and physiological fluctuation of effusion volume within the 30-day CBCT-to-MRI window cannot be fully excluded, so residual label noise may remain.
