Abstract
Knee osteoarthritis (OA) severity is graded from plain radiographs using the Kellgren–Lawrence (KL) system, but classifier errors are concentrated at the early adjacent-grade boundaries KL0 vs. KL1 and KL1 vs. KL2, where radiographic differences are subtle and reader labels are ambiguous. We ask whether boundary-focused generative augmentation can improve classification at these difficult transitions, and we compare four augmentation paradigms under matched conditions: an unconditional GAN, a class-conditional GAN, a counterfactual GAN (CycleGAN-based), and a class-conditional diffusion model. Five backbones (Siamese, EfficientNet-B5, ConvNeXt, ConvNeXt V2, MaxViT) are trained on Osteoarthritis Initiative (OAI) radiographs after percentile-based filtering of cases near the decision boundary, and synthetic and real images are combined at five real-to-synthetic ratios. Beyond internal evaluation, we test best-per-strategy models on two independent external cohorts. On KL0 vs. KL1, the counterfactual GAN produced the largest internal gains (balanced accuracy improvements of 0.05–0.08 over augmented baselines; p < 0.01, paired bootstrap), with the conditional diffusion model close behind. On KL1 vs. KL2, unconditional and counterfactual augmentations were most stable internally, while the conditional diffusion model gave the strongest external generalization on one cohort. A 5:1 real-to-synthetic ratio and the narrower 35–65% boundary filter were consistently preferable. Boundary-focused generative augmentation can improve adjacent-grade KL classification, but the benefit is paradigm, backbone, and cohort-dependent.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RHEUM_AI_002.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=F1NgssPBGL
BibTex
@InProceedings{ShaAhm_Augmenting_MICCAISAT2026,
author = { Shamsia, Ahmed AND Pham, Ngoc Huy AND Saarakkala, Simo AND Bayramoglu, Neslihan},
title = { { Augmenting the Grey Zone: Generative Models for Ambiguous Kellgren–Lawrence Grade Boundaries in Knee Osteoarthritis } },
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}
}