Using Unlabeled Data in Self-Supervised Training Improves Automatic Vertebral Body Fracture Assessment

Vertebral body (VB) fractures can have severe implications for the patient’s well-being, but often remain undetected. Automatic vertebral fracture assessment is challenging due to severe class imbalance in fracture grades, scarcity of annotated data, and often subtle visual differences between the grades. In this paper, we show that leveraging unlabeled data, collected from eight diverse, publicly available datasets in a self-supervised way via the BYOL framework, can noticeably improve performance. Our pre-trained models outperformed the non-pre-trained baselines in 17 out of 18 comparisons, 11 of which were statistically significant, when evaluated on the widely-used VerSe dataset, our in-house dataset, and combinations thereof. Furthermore, our models reached an average AUROC value of 99.6% on the VerSe test corpus, the highest recorded in the literature so far.