NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Self-Supervised Vision Transformers for CBCT-Based Detection of Temporomandibular Joint Osteoarthritis
TL;DR
Temporomandibular joint osteoarthritis (TMJ OA) is difficult to detect using cone-beam CT (CBCT) due to subtle changes. A pipeline using self-supervised Vision Transformers, specifically DINOv2 with partial unfreezing, was developed for patient-level classification of OA versus normal.
✦ Why It Matters
Engineers can leverage these findings to improve automated detection systems in medical imaging with limited data.
Key Takeaways
- Explore self-supervised models for medical imaging tasks.
- Consider partial unfreezing of transformer layers for better performance.
- Utilize attention-based multiple instance learning for classification.
- Focus on adaptation strategies over model choice in low-data scenarios.
Full Summary
Temporomandibular joint osteoarthritis (TMJ OA) is a common degenerative condition that poses challenges for automated detection due to subtle osseous changes visible on cone-beam CT (CBCT). Researchers explored the effectiveness of the DINO family of self-supervised Vision Transformers, including DINOv1 and DINOv2, in adapting to CBCT data.
They implemented a slice-based pipeline where axial CBCT slices were encoded using a Vision Transformer (ViT) and aggregated through attention-based multiple instance learning (MIL) for binary classification. Systematic experiments revealed that partially unfreezing the last two transformer blocks significantly enhanced performance, raising the AUC from 0.671 with a fully frozen DINOv2 to 0.902.
This result surpassed other models, including DINOv1 and a supervised ImageNet baseline. The findings suggest that the choice of adaptation strategy is more critical than the backbone model in low-data medical imaging contexts, providing practical guidance for future applications.