Uncertainty-Driven Self-Supervised Learning with Test-Time Adaptation for Anomaly Detection of T2 Hyperintensity in Spinal Cord

Motivation: Detecting T2 hyperintensities in cervical spinal MRI is essential for diagnosing cervical spondylotic myelopathy, but current methods struggle with sensitivity and specificity in complex regions. Goal(s): To develop a self-supervised anomaly detection framework that leverages uncertainty to enhance diagnostic accuracy. Approach: Our Masked Autoencoder-based model uses epistemic uncertainty to focus on challenging normal regions and aleatoric uncertainty to prevent overfitting in anomalous areas. A two-stage, uncertainty-driven adaptation maximizes reconstruction accuracy without manual annotations. Results: The model achieved top performance across classification (F1 = 0.8336) and localization (F1 = 0.7345) tasks, surpassing supervised and unsupervised benchmarks. Impact: This framework advances automated anomaly detection in spinal MRI, providing clinicians with accurate, quantifiable, and localized metrics for T2 hyperintensities. Future researches will further utilize the uncertainty estimation to improve the anomaly detection performance in other diseases.