Bidirectional generative modeling of knee osteoarthritis: pseudo-healthy and progressive radiograph synthesis using conditional GANs and GAN-based classification
Document Type
Article
Publication Date
8-15-2026
Abstract
Knee osteoarthritis (KOA) is a progressive joint disease and a leading cause of disability worldwide. Its severity is typically assessed using the Kellgren–Lawrence (KL) grading system, but manual grading is subjective and often struggles to detect early disease. Automated approaches using deep learning have shown promise, yet progress is limited by dataset scarcity, labeling noise, and the inability of conventional classifiers to capture the temporal nature of osteoarthritis progression. In this study, we present a bidirectional generative adversarial framework capable of transforming radiographs between healthy/mild (KL01) and moderate/severe (KL234) states. The generator produces both pseudo-healthy regressions (KL234 → KL01) and synthetic progressions (KL01 → KL234), while the discriminator is repurposed as a classifier through adaptive thresholding. This dual capability enables both the synthesis of disease trajectories and robust binary classification without the need for a separate supervised network. Experiments on a large dataset of 320,000 radiographs show that the discriminator achieves classification accuracies approaching 99% with an AUC exceeding 0.95. Qualitative analysis further indicates that generated images capture clinically relevant features, including joint space narrowing and osteophyte formation. Saliency mapping and t-SNE visualizations confirmed that the model emphasizes radiographic hallmarks and aligns synthetic images along a plausible disease trajectory. These findings highlight the potential of bidirectional GANs not only as augmentation tools but also as a means of modeling disease progression and regression in KOA. With further refinement and expert validation, such frameworks could contribute to diagnostic support, prognostic modeling, and educational applications in musculoskeletal radiology.
Keywords
Deep learning, Discriminator-based classification, Generative adversarial network, Knee osteoarthritis, Medical image translation, Predictive image synthesis, Radiographic analysis
Publication Title
Biomedical Signal Processing and Control
ISSN
1746-8094
DOI
10.1016/j.bspc.2026.110372
Recommended Citation
Ng, Keng Seng; Goh, Siew Li; Hasikin, Khairunnisa; and Lai, Khin Wee, "Bidirectional generative modeling of knee osteoarthritis: pseudo-healthy and progressive radiograph synthesis using conditional GANs and GAN-based classification" (2026). Research Publications (2026 to 2030). 500.
https://knova.um.edu.my/research_publications_2026_2030/500
Volume
122