arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Signal ProcessingEnglishPublished

Routine CT scans can spot hidden joint fluid without waiting for MRIs

November 19, 2026arXiv: 2611.23042v1

A new deep‑learning method can detect soft‑tissue swelling in the temporomandibular joint (TMJ) from ordinary cone‑beam computed tomography (CBCT) scans. The technique identifies fluid accumulation that is normally invisible to the naked eye on CT, yet confirmed by corresponding magnetic‑resonance imaging (MRI).

The system uses a frozen 3‑D masked autoencoder (BertsWin) that was pre‑trained on 6 550 unlabelled CBCT volumes. During training the network focuses on the soft‑tissue window (–300 to 300 HU) through a physics‑aware loss (PhysLoss) and a GradientConductor (GCond) optimizer. It then localises the condylar zenith of the jaw with an Attention‑U‑Net head and crops a tight 8 × 8 × 8‑token box, removing 81 % of bone background.

Across 30 patient‑isolated, 3 × 10‑fold cross‑validation splits the frozen encoder achieves an area‑under‑the‑curve of 0.793 ± 0.036 and a Youden’s J of 0.459. Performance falls to around chance when the same architecture is pre‑trained with a standard L2 loss or with unmasked variance decomposition (AUC ≈ 0.5). These results highlight that only the coupled PhysLoss + GCond pre‑training preserves true 3‑D anatomical cues.

To verify that the network really focuses on intra‑articular tissue, Grad‑CAM heat maps were produced for the validation cohort. The maps (Fig. 2) show peak activation confined to the joint cavity, confirming that the model uses physiologic features rather than spurious background patterns.

The study uses a small cohort: 212 paired CBCT–MRI joints from 99 patients, all scanned within 30 days of each other. Such sampling limits statistical power for rare conditions and may add label noise when fluid volumes change between scans.

Overall, the approach demonstrates that routine, low‑contrast CT can be opportunistically repurposed for soft‑tissue disease screening, potentially expanding diagnostic coverage without extra imaging.

Figure 2: Grad‑CAM attention maps for cross‑modal synovial effusion detection across axial, coronal, and sagittal CBCT views. Yellow boxes mark the cropped region; the white crosshair marks peak activation inside the joint.

Source paper

Uncovering the Invisible in Routine Radiology: Cross-Modal Detection of MRI-Verified Soft-Tissue Pathologies from CBCT via Physics-Guided 3D Self-Supervised Learning

Signal ProcessingComputer VisionMachine Learning
View on arXiv
Routine CT scans can spot hidden joint fluid without waiting for MRIs | arXiv News