ABLSpineLevelCheck: Localization of Vertebral Levels on Fluoroscopy via Semi-supervised Abductive Learning

Deep learning has demonstrated promising efficacy in the localization of vertebrae within X-ray images, although it is recognized for its deficiencies in compositional generalization, data efficiency, and interpretability. To address this issue, we introduce an abductive learning mechanism, situated within the neuro-symbolic paradigm, tailored for semi-supervised vertebral localization. Initially, unannotated spinal fluoroscopic images are processed by the networks to infer pseudo-labels for vertebra localization. Subsequently, these pseudo-labels undergo abductive reasoning via a knowledge base comprised of first-order logical clauses. The networks are then re-trained using the abducted outcomes. Additionally, we propose an ensemble technique that amalgamates semantic detection of vertebral levels with instance detection. To further enhance performance, we have synthesized a dataset and annotated the BUU dataset for network pre-training. Ablation studies validate the efficacy of the proposed components in our methodology. Furthermore, comparative analyses reveal that our approach significantly surpasses the leading object detection algorithms, exhibiting superior performance with minimal annotations.