Abstract
Knee osteoarthritis (KOA) and calcium pyrophosphate deposition disease (CPPD) are distinct but clinically related conditions that can be assessed from the same knee radiograph using ordinal severity scales, namely the Kellgren–Lawrence (KL) scale for KOA and a dedicated CPPD severity scale. Automated grading is complicated by subtle differences between adjacent grades, class imbalance, and the asymmetric relationship between KL and CPPD severity. This study investigates whether the two diseases are better predicted using a shared visual representation or separate task-specific encoders. We adapt DINOv2-S and DINOv3-S foundation models through Low-Rank Adaptation (LoRA) or full fine-tuning and compare shared multi-head models with disjoint single-task models under nominal and ordinal soft-label supervision. Experiments are conducted on 965 radiographs with paired KL and CPPD annotations. The best overall configuration is a fully fine-tuned DINOv2-S shared encoder trained with beta-shaped ordinal soft labels, achieving a mean quadratic weighted kappa of 0.844 across the two tasks, with task-specific values of 0.830 for KL and 0.857 for CPPD. It also obtains the lowest mean MAE and AMAE among the evaluated configurations. Full fine-tuning generally outperforms LoRA in the multi-task setting, while the effect of ordinal supervision depends on the backbone and adaptation strategy. Although task-specific models remain competitive for CPPD alone, sharing the encoder provides the best overall trade-off for simultaneous KL and CPPD grading.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RHEUM_AI_008.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=gATFdLBhWo
BibTex
@InProceedings{DiAle_Towards_MICCAISAT2026,
author = { Di Matteo, Alessandro AND Fiorentino, Maria Chiara AND Rosati, Riccardo AND Moccia, Sara},
title = { { Towards Multi-Task Ordinal Grading of Knee Osteoarthritis and Calcium Pyrophosphate Deposition Using Vision Foundation Models } },
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}
}