EVALUATING UNSUPERVISED ANOMALY DETECTION FOR AUTOMATED SHOMRI IN A YOUNG ADULTS POPULATION

Early structural changes in hip OA are rarely evaluated in young adults, yet identifying such changes may provide insight into disease onset. The Scoring Hip Osteoarthritis with MRI (SHOMRI) system enables detailed assessment of hip joint pathology, but large-scale application is limited by the time and expertise required for manual scoring. Because abnormalities in young adults are rare and subtle, unsupervised anomaly detection (UAD), which learns only from normal examples, may offer a scalable approach for automated screening.

The aim of this study was to investigate whether UAD models can prescreen SHOMRI scoring in young adults by automatically identifying potential abnormalities on MRI.

Hip MRI from 286 participants (572 hips; 145 males, 141 females; mean age 18.3 ± 0.7 years) from the Generation R cohort were manually scored using SHOMRI, which evaluates 8 features across 10 hip joint regions. Scoring was performed on T2-weighted fat-suppressed 3D FSE (CUBE) images (0.89 × 0.89 mm² inplane resolution; 1.2 mm slice thickness). Four reconstruction-based UAD models: an autoencoder (AE), a variational autoencoder (VAE), a generative adversarial network (GAN), and diffusion models were trained using 512 healthy hips and evaluated on 60 hips with and without abnormalities. Model performance was assessed through visual inspection of reconstructions, separation of anomaly score distributions using histograms, and classification performance using AUC.

Only 29 hips (5%) showed SHOMRI abnormalities, mostly (n=26) labral changes (Table 1). All evaluated UAD models demonstrated poor discrimination, with overlapping anomaly score distributions. Reconstructions were blurry and lacked fine anatomical detail: abnormalities were suppressed, and normal variation smoothed, producing anomaly maps mostly highlighting intensity differences. AE and VAE models flagged nearly all cases as abnormal, GANs introduced hallucinated edges, and diffusion models suppressed both normal and abnormal features.

Structural abnormalities were rare in this young adult population. The small size and subtle appearance of SHOMRI findings relative to normal anatomic variability, combined with low‑fidelity reconstructions, limited UAD performance. Methods incorporating anatomical priors, region‑specific modeling, or weak supervision may be needed to improve detection of subtle SHOMRI abnormalities.