Abstract
Bone erosions are a key marker of joint damage in rheumatoid arthritis (RA), yet segmenting them on hand radiographs remains very difficult, with reported full-hand dense-segmentation performance on the public RAM-H1200 benchmark staying below 20% Dice. This paper analyzes the factors behind this performance ceiling. Erosions are tiny, covering roughly 0.04% of the image, and lie along the bone margin, where they closely resemble normal anatomy, causing models to produce frequent false positives on healthy joints. We first show that healthy and eroded joint crops are barely separable from crop-level features alone. We then evaluate strategies to add information, including a cortical-margin prior and a healthy-anatomy residual. Under a matched operating point, these variants settle onto essentially the same sensitivity/false-positive curve as a plain U-Net baseline. A contralateral comparison adds nothing over a joint-matched healthy prototype. These results consistently hold across independent training seeds and replicate on the held-out test set. Reframing the task as patch-level detection yields a deceptively high patch ROC-AUC (≈0.84); when evaluated per joint, performance drops to 0.59–0.61, below joint-level segmentation (≈0.64). Using larger 512 × 512 crops, which retain more of the original radiograph resolution, also yields no improvement. Ultimately, the performance bottleneck lies in the isolated joint view itself, rather than model architecture, conditioning, task formulation, or image resolution. Breaking this ceiling will require context beyond single-joint imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RHEUM_AI_007.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=lwZF5UJgd6
BibTex
@InProceedings{TanArt_Why_MICCAISAT2026,
author = { Taneja, Arti AND Omer, Hasmila A. AND Vasikarla, Shantaram},
title = { { Why Isolated Joint Views Limit Bone-Erosion Segmentation on Hand Radiographs } },
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}
}