pith. sign in

arxiv: 2606.08364 · v1 · pith:TDVXPEZMnew · submitted 2026-06-06 · 💻 cs.CV · cs.AI

Self-Supervised Vision Transformers for CBCT-Based Detection of Temporomandibular Joint Osteoarthritis

Shradhdha Trivedi , Vrundan Sojitra , Mariela Padilla This is my paper

Pith reviewed 2026-06-27 19:35 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords self-supervised vision transformersDINOCBCTTMJ osteoarthritismultiple instance learningtransfer learningpartial fine-tuningmedical image classification
0
0 comments X

The pith

Partial unfreezing of the final two transformer blocks in DINOv2 raises AUC from 0.671 to 0.902 for TMJ osteoarthritis detection on CBCT.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests how DINO-family self-supervised vision transformers transfer to cone-beam CT for binary classification of temporomandibular joint osteoarthritis. It evaluates different unfreezing strategies on the ViT backbone within a per-slice encoding pipeline that aggregates via attention-based multiple instance learning. Systematic ablations show that unfreezing only the last two blocks drives the largest gain and beats both other DINO variants and a supervised ImageNet baseline. A reader would care because the work isolates a minimal adaptation recipe that works in low-data medical settings where full retraining is impractical.

Core claim

Through ablations across unfreezing strategies and aggregation designs on a multi-source CBCT dataset, partial unfreezing of the final two transformer blocks is the decisive factor, improving AUC from 0.671 with fully frozen DINOv2 to 0.902. This result outperforms DINOv1 at 0.867, DINOv2+reg at 0.774, and a supervised ImageNet ViT-B/16 baseline at 0.843 in patient-level OA versus normal classification.

What carries the argument

Selective unfreezing of only the final two transformer blocks in the ViT backbone, paired with attention-based multiple instance learning that aggregates per-slice encodings into a patient-level decision.

If this is right

  • Adaptation strategy matters more than backbone choice among DINO variants for this task.
  • Partial unfreezing enables effective transfer without requiring large medical pretraining corpora.
  • Attention MIL aggregation of 2D slices produces usable patient-level labels under the tested conditions.
  • The performance ordering holds across multiple aggregation designs in the reported ablations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same minimal unfreezing recipe may transfer to other low-data radiology tasks that rely on subtle osseous changes.
  • Direct comparison against 3D-native architectures on the same data would test whether slice aggregation loses critical volumetric context.
  • Prioritizing fine-tuning depth over pretraining source could become a default first step when adapting foundation models to new medical modalities.

Load-bearing premise

The multi-source CBCT dataset captures clinical variability and the slice-based 2D processing with attention MIL preserves the essential 3D pathology information.

What would settle it

An independent test set where fully frozen backbones or full 3D volumetric models achieve equal or higher AUC than the partial-unfreezing pipeline would falsify the claim that this adaptation step is decisive.

Figures

Figures reproduced from arXiv: 2606.08364 by Mariela Padilla, Shradhdha Trivedi, Vrundan Sojitra.

Figure 1
Figure 1. Figure 1: Proposed pipeline. (1) A 3D CBCT volume (H ×W ×D) is input. (2) N axial slices are extracted from the TMJ region. (3) Each slice is independently encoded by a shared pretrained DINO-family ViT backbone (weights partially adapted). (4) Per-slice [CLS] embeddings (N ×D) are formed. (5) Attention-based MIL aggregation learns to up-weight diagnostically informative slices. (6) A linear classifier outputs binar… view at source ↗
read the original abstract

Temporomandibular joint osteoarthritis (TMJ OA) is a prevalent degenerative condition whose osseous changes are often subtle on cone-beam CT (CBCT), making automated detection challenging. We study how well the DINO family of self-supervised vision transformers -- DINOv1, DINOv2, DINOv2+reg, and RAD-DINO (a radiology-pretrained variant) -- transfers to CBCT, asking how much backbone adaptation is needed and of what kind. We propose a simple slice-based pipeline using Vision Transformer (ViT) backbones: axial CBCT slices are encoded per-slice by a frozen or partially adapted ViT and aggregated via attention-based multiple instance learning (MIL) for patient-level binary OA/Normal classification. Through systematic ablation across unfreezing strategies and aggregation designs on a multi-source CBCT dataset, we find that partial unfreezing of the final two transformer blocks is the decisive factor, improving AUC from 0.671 (fully frozen DINOv2) to 0.902. This outperforms DINOv1 (0.867), DINOv2+reg (0.774), and a supervised ImageNet ViT-B/16 baseline (0.843). Our results provide practical guidance for adapting DINO-family foundation models in low-data medical imaging settings, showing that adaptation strategy is a stronger driver of performance than backbone choice alone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper claims that DINO-family self-supervised ViTs can be effectively adapted for CBCT-based TMJ OA detection via a slice-based pipeline (axial slices encoded by ViT, aggregated by attention MIL). Systematic ablations show that partial unfreezing of the final two transformer blocks in DINOv2 is the decisive factor, raising AUC from 0.671 (fully frozen) to 0.902 and outperforming DINOv1 (0.867), DINOv2+reg (0.774), and a supervised ImageNet ViT-B/16 baseline (0.843). The work positions adaptation strategy as more important than backbone choice for low-data medical imaging.

Significance. If the empirical results hold under proper validation, the paper supplies concrete, actionable guidance on efficient adaptation of self-supervised foundation models in radiology, demonstrating that targeted unfreezing can yield large gains without full fine-tuning. This is particularly relevant for data-scarce domains and could influence practical deployment of ViT-based pipelines.

major comments (2)
  1. [Abstract] Abstract: the central claim that partial unfreezing of the final two blocks drives the AUC gain (0.671 → 0.902) rests on the untested assumption that axial-slice 2D encoding plus attention MIL fully captures 3D osseous changes (e.g., condylar morphology across slices). No 3D volumetric baseline, slice-count statistics, or inter-slice spatial analysis is reported to rule out information loss as a confound.
  2. [Abstract] Abstract: performance numbers are given without dataset size, number of patients/slices, cross-validation folds, inter-rater variability, or statistical significance tests (e.g., DeLong test for AUC differences). These omissions make it impossible to verify whether the reported improvements are robust or dataset-specific.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and have made targeted revisions to the abstract and discussion to improve clarity and completeness.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that partial unfreezing of the final two blocks drives the AUC gain (0.671 → 0.902) rests on the untested assumption that axial-slice 2D encoding plus attention MIL fully captures 3D osseous changes (e.g., condylar morphology across slices). No 3D volumetric baseline, slice-count statistics, or inter-slice spatial analysis is reported to rule out information loss as a confound.

    Authors: We acknowledge that our pipeline is explicitly 2D slice-based with attention MIL for aggregation, which is a deliberate design choice for computational efficiency in CBCT analysis. The manuscript does not include a 3D volumetric baseline, as the study scope centers on adaptation strategies for 2D self-supervised ViTs rather than comparing 2D vs. 3D architectures. We have added text to the abstract and a dedicated limitations paragraph in the discussion noting this scope and the potential value of future 3D comparisons. Slice-count statistics and basic inter-slice aggregation details are already present in the Methods section. revision: partial

  2. Referee: [Abstract] Abstract: performance numbers are given without dataset size, number of patients/slices, cross-validation folds, inter-rater variability, or statistical significance tests (e.g., DeLong test for AUC differences). These omissions make it impossible to verify whether the reported improvements are robust or dataset-specific.

    Authors: We agree these details strengthen verifiability. While the full manuscript already reports dataset composition, patient/slice counts, 5-fold cross-validation, and statistical testing (including DeLong tests for AUC comparisons) in the Methods and Results, the abstract omitted a concise summary. We have revised the abstract to include these elements along with a note on expert annotation. Inter-rater aspects are covered in the supplementary material. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical ablation on held-out data

full rationale

The paper is an empirical study that encodes axial CBCT slices with DINO-family ViTs (frozen or partially unfrozen), aggregates via attention MIL, and reports AUC on held-out patient-level labels. All performance numbers (e.g., 0.671 frozen vs. 0.902 with last two blocks unfrozen) are measured outcomes on independent test splits; no equations, fitted parameters renamed as predictions, or self-citation chains are invoked to derive the central claim. The choice of which blocks to unfreeze is presented as an experimental result, not a mathematical necessity. No load-bearing step reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on experimental results from a specific dataset and model adaptations rather than new axioms or entities; the main inputs are pre-trained DINO models and the CBCT data.

axioms (1)
  • domain assumption Self-supervised pre-trained ViTs can be effectively adapted to medical imaging domains by partial fine-tuning.
    The paper's pipeline relies on this transfer learning assumption without proving it from first principles.

pith-pipeline@v0.9.1-grok · 5793 in / 1317 out tokens · 30384 ms · 2026-06-27T19:35:20.387773+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Eric Schiffman, Richard Ohrbach, Edmond Truelove, John Look, Gary Anderson, Jean-Paul Goulet, Thomas List, Peter Svensson, Yoly Gonzalez, Frank Lobbezoo, et al. Diagnostic criteria for tempo- romandibular disorders (DC/TMD) for clinical and research applica- tions: Recommendations of the international RDC/TMD consortium network and orofacial pain special ...

    2014

  2. [2]

    Artificial intelligence for detecting temporomandibular joint osteoarthritis on imaging: A systematic review and meta-analysis.PLOS ONE, 18(2):e0280562, 2023

    Lingling Xu, Qasim Al-Dwairi, Yuxin Chen, Yanpeng Cao, Xiao- jie Zheng, Shanyong Wu, and Fang Hua. Artificial intelligence for detecting temporomandibular joint osteoarthritis on imaging: A systematic review and meta-analysis.PLOS ONE, 18(2):e0280562, 2023

    2023

  3. [3]

    Kang-Il Lee, Jae-Joon Hwang, Yoon-Ji Choi, In-Young Park, Ho- Gul Jeong, and Sang-Sun Han. Evaluation of artificial intelligence for detecting temporomandibular joint osteoarthritis on panoramic radiography.Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology, 130(2):e58, 2020

    2020

  4. [4]

    Deep learning for osteoarthritis classification in the temporomandibular joint.Oral Diseases, 29(3):1050–1059, 2023

    Woosung Jung, Kyung-Eun Lee, Bong-Jae Suh, Hyun Seok, and Dong-Won Lee. Deep learning for osteoarthritis classification in the temporomandibular joint.Oral Diseases, 29(3):1050–1059, 2023

    2023

  5. [5]

    Gulsun Eser, Mevlut Kara, and Akin Ozkan. Artificial intelligence- based automatic detection and classification of temporomandibular joint osteoarthritis on cone-beam computed tomography.Journal of Oral Rehabilitation, 50(9):863–870, 2023

    2023

  6. [6]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650– 9660, 2021

    2021

  7. [7]

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galvez, Russell Baqu´e, Priya Bhargava, Quentin Duval, Rodrigo Jenatton, Armand Joulin, Ishan Misra, Gabriel Syn- naeve, Camille Couprie, Herv ´e J´...

    2024

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  9. [9]

    Vision Transformers Need Registers

    Timoth´ee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bo- janowski. Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Jaeger, Hans-Ulrich Kauczor, and Klaus H

    Daniel Wolf, Julian R¨uckert, Paul F. Jaeger, Hans-Ulrich Kauczor, and Klaus H. Maier-Hein. Self-supervised pretraining for 2D medical image segmentation.Scientific Reports, 13:9379, 2023

    2023

  11. [11]

    Tomczak, and Max Welling

    Maximilian Ilse, Jakub M. Tomczak, and Max Welling. Attention- based deep multiple instance learning. InProceedings of the 35th International Conference on Machine Learning (ICML), pages 2127– 2136, 2018

    2018

  12. [12]

    Chen, Chengkuan Chen, Yicong Li, Tiffany Y

    Richard J. Chen, Chengkuan Chen, Yicong Li, Tiffany Y . Chen, An- drew D. Trister, Nicholas J. Durr, and Faisal Mahmood. Scaling vi- sion transformers to gigapixel images via hierarchical self-supervised learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16144–16155, 2022

    2022

  13. [13]

    Castro, Anton Schwaighofer, Aditya Nori, Javier Alvarez-Valle, Ozan Oktay, and Noel Codella

    Fernando P´erez-Garc´ıa, Harshita Sharma, Sam Bond-Taylor, Kenza Bhatt, Valentina Salvatelli, Maximilian Ilse, Shruthi Bannur, Osman Bouzid, Daniel C. Castro, Anton Schwaighofer, Aditya Nori, Javier Alvarez-Valle, Ozan Oktay, and Noel Codella. RAD-DINO: Explor- ing scalable medical image encoders beyond text supervision.arXiv preprint arXiv:2401.10815, 2024

    work page arXiv 2024

  14. [14]

    A comprehensive patient-specific prediction model for temporomandibular joint osteoarthritis progression

    Lucia Cevidanes, Najla Al Turkestani, Jonas Bianchi, Marilia Gurgel, and Tengfei Li. A comprehensive patient-specific prediction model for temporomandibular joint osteoarthritis progression. University of Michigan Deep Blue Data Repository, https://doi.org/10. 7302/xc32-4d53, 2024

    2024