COMPARISON OF THE PERFORMANCE OF PRETRAINED DEEP LEARNING MODELS FOR THE AUTOMATIC KELLGREN-LAWRENCE GRADING OF KNEE OSTEOARTHRITIS USING PLAIN RADIOGRAPHS
Bozok Tıp Dergisi · 15.03.2026 · 16 (1) · 115-125
DOI: 10.16919/bozoktip.1859321 ↗Kurz gefasst
Fachartikel in der Zeitschrift Bozok Tıp Dergisi, online erschienen am 15.03.2026. Verfasst von Hafize Kızılkaya und 2 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Gefördert von: Yozgat Bozok University.
Abstract (Originalsprache)
Objective
Radiographic assessment of knee osteoarthritis (OA) commonly relies on the Kellgren–Lawrence (KL) grading system; however, its subjective nature leads to considerable inter- and intra-observer variability, particularly in early disease stages. This study aimed to comparatively evaluate pre-trained deep learning models for automated KL grading of knee OA from plain radiographs using an ordinal-aware learning and evaluation framework. Materials and Methods: This retrospective experimental study utilized 8,260 knee radiographs obtained from the publicly available Osteoarthritis Initiative (OAI) dataset, with expert-assigned KL grades ranging from 0 to 4. Five pre-trained convolutional neural network architectures (VGG-16, ResNet-50, DenseNet-121, EfficientNetB0, and InceptionV3) were implemented using transfer learning. All models were trained under identical preprocessing, augmentation, class-balancing, and hyperparameter settings to ensure fair comparison. An ordinal CORAL-based loss function was employed to model the ordered nature of KL grades. Model performance was primarily evaluated using quadratic weighted kappa (QWK), along with accuracy, balanced accuracy, macro-F1 score, ROC–AUC, and precision–recall analyses. Decision curve analysis (DCA) was conducted at clinically relevant thresholds (KL ≥ 2 and KL ≥ 3) to assess potential clinical utility.
Results
Among the evaluated architectures, VGG-16 achieved the highest ordinal agreement on the independent test set (QWK = 0.830), with a macro-F1 score of 0.676 and balanced accuracy of 0.684. Overall, model performance was higher for moderate-to-severe OA stages (KL grades 3 and 4), while lower discriminative performance was observed for early-stage disease, particularly KL grade 1. Confusion matrix analysis demonstrated that most misclassifications occurred between adjacent KL grades, indicating clinically plausible ordinal behavior. Decision curve analysis revealed that the proposed ordinal deep learning model provided a consistently higher net benefit than treat-all and treat-none strategies across a wide range of threshold probabilities for both KL ≥ 2 and KL ≥ 3 scenarios.
Conclusion
Ordinal-aware deep learning models can effectively perform automated KL grading of knee osteoarthritis from plain radiographs, yielding clinically meaningful and interpretable results. The proposed framework reduces observer-dependent variability and demonstrates potential as a decision-support tool for both early and advanced stages of knee OA. Further validation using multi-center datasets is warranted to enhance clinical generalizability.
Förderung
- Yozgat Bozok University
Zitieren
APA
Kızılkaya, H., Ortataş, F. N., & Üreten, K. (2026). COMPARISON OF THE PERFORMANCE OF PRETRAINED DEEP LEARNING MODELS FOR THE AUTOMATIC KELLGREN-LAWRENCE GRADING OF KNEE OSTEOARTHRITIS USING PLAIN RADIOGRAPHS. Bozok Tıp Dergisi, 16(1), 115-125. https://doi.org/10.16919/bozoktip.1859321
Vancouver
Kızılkaya H, Ortataş FN, Üreten K. COMPARISON OF THE PERFORMANCE OF PRETRAINED DEEP LEARNING MODELS FOR THE AUTOMATIC KELLGREN-LAWRENCE GRADING OF KNEE OSTEOARTHRITIS USING PLAIN RADIOGRAPHS. Bozok Tıp Dergisi. 2026;16(1):115-125. doi:10.16919/bozoktip.1859321
BibTeX
@article{Kzlkaya2026,
author = {Kızılkaya, Hafize and Ortataş, Fatma Nur and Üreten, Kemal},
title = {COMPARISON OF THE PERFORMANCE OF PRETRAINED DEEP LEARNING MODELS FOR THE AUTOMATIC KELLGREN-LAWRENCE GRADING OF KNEE OSTEOARTHRITIS USING PLAIN RADIOGRAPHS},
journal = {Bozok Tıp Dergisi},
year = {2026},
volume = {16},
number = {1},
pages = {115--125},
doi = {10.16919/bozoktip.1859321}
}
EUVIMED