Abstract
Manual Sharp–van der Heijde (SvdH) scoring requires expert readers to grade dozens of individual joints per radiograph, making it slow and variable between readers. We ask which ingredients actually matter when transferring a foundation model to this task. On the RAM-H1200 benchmark, partially fine-tuned DINOv2 ViT-S/14 shows that generic self-supervised features alone already surpass the strongest published supervised baselines for both bone erosion (BE) and joint-space narrowing (JSN). The largest further gain comes not from domain knowledge but from the objective: replacing cross-entropy with a cumulative ordinal loss raises quadratic weighted kappa to 0.591 ± 0.028 for BE and 0.634 ± 0.006 for JSN, and the advantage widens as labels are removed, reaching +0.069 QWK for BE at a quarter of the training data. We then test three domain-specific extensions that should plausibly help: continual self-supervised pretraining on 30,000 unlabelled multicentre hand and wrist radiographs, explicit joint-type conditioning and BE/JSN multi-task learning. Code: https://github.com/EIDOSLAB/OrdinalSvdH.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RHEUM_AI_005.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/RHEUM_AI_005_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=0UjW88E2El
BibTex
@InProceedings{CanMic_Ordinal_MICCAISAT2026,
author = { Cannito, Michele AND Renzulli, Riccardo AND Grangetto, Marco},
title = { { Ordinal DINOv2 Transfer for Joint-Level SvdH Scoring: What Helps and What Does Not } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17271},
month = {pending},
page = {pending}
}