An Intrinsically Explainable Approach to Detecting Vertebral Compression Fractures in CT Scans via Neurosymbolic Modeling

TL;DR

This work tackles vertebral compression fracture detection in CT scans for opportunistic screening with a focus on interpretability. It proposes a neurosymbolic framework that fuses DL-based vertebral segmentation with a shape-based height-map analysis, and derives a sparse, rule-based decision via RuleFit, yielding the interpretable predictor $\hat{f}(x) = \hat{\beta}_0 + \sum_{k=1}^K {\hat{\alpha_k} r_k(x)}$. On the VerSe19 dataset, the method achieves $\text{accuracy}=0.96$ and $\text{sensitivity}=0.91$, matching or exceeding a DenseNet baseline while providing intuitive, rule-grounded explanations for each prediction. This intrinsic transparency can enhance clinician trust and support informed diagnosis and treatment planning for VCFs. Future work will explore trust calibration and usability to maximize adoption in clinical workflows.

Abstract

Vertebral compression fractures (VCFs) are a common and potentially serious consequence of osteoporosis. Yet, they often remain undiagnosed. Opportunistic screening, which involves automated analysis of medical imaging data acquired primarily for other purposes, is a cost-effective method to identify undiagnosed VCFs. In high-stakes scenarios like opportunistic medical diagnosis, model interpretability is a key factor for the adoption of AI recommendations. Rule-based methods are inherently explainable and closely align with clinical guidelines, but they are not immediately applicable to high-dimensional data such as CT scans. To address this gap, we introduce a neurosymbolic approach for VCF detection in CT volumes. The proposed model combines deep learning (DL) for vertebral segmentation with a shape-based algorithm (SBA) that analyzes vertebral height distributions in salient anatomical regions. This allows for the definition of a rule set over the height distributions to detect VCFs. Evaluation of VerSe19 dataset shows that our method achieves an accuracy of 96% and a sensitivity of 91% in VCF detection. In comparison, a black box model, DenseNet, achieved an accuracy of 95% and sensitivity of 91% in the same dataset. Our results demonstrate that our intrinsically explainable approach can match or surpass the performance of black box deep neural networks while providing additional insights into why a prediction was made. This transparency can enhance clinician's trust thus, supporting more informed decision-making in VCF diagnosis and treatment planning.

Figures (2)

• Figure 1: \ref{['fig:overview']} The height map extraction process relies on vertebral body segmentation with TotalSegmentator wasserthal2023totalsegmentator to obtain height maps in a consistent local coordinate system. \ref{['fig:sections']} Drawing on real diagnostic decision-making, we define 7 regions of the vertebral body to inform the severity rating.
• Figure 2: \ref{['fig:visComp']} Our neurosymbolic reasoning strategy evaluates the relative height of the vertebral body in specific regions to identify cases of VCF. \ref{['fig:CM']} Example model output. A common failure mode occurs when the vertebral body's shape suggests a compression fracture, even if one is not labeled.