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

Yiqing Shen, Blanca Inigo Romillo, Yiqing Shen 0003, Benjamin D. Killeen, Michelle Song, Axel Krieger, Christopher Bradley, Mathias UnberatharXiv 2024 · DOI · arXiv

Introduction

Vertebral compression fractures (VCFs) are a hallmark of osteoporosis, yet they are notoriously underdiagnosed. Often, these fractures are present in CT scans performed for unrelated reasons, such as chest or abdominal imaging. "Opportunistic screening"—the use of AI to automatically flag these fractures—represents a significant opportunity to improve patient outcomes and reduce healthcare costs.

However, clinical adoption of AI is frequently hindered by the "black-box" nature of state-of-the-art deep learning models. While these models may achieve high accuracy, they often fail to provide human-readable justifications for their predictions. Unreliable post-hoc explanation techniques, such as saliency maps, further erode trust in high-stakes medical environments. This research introduces a neurosymbolic AI approach that is "intrinsically explainable," combining deep learning's perceptual capabilities with a transparent, rule-based logic system to detect VCFs with 96% accuracy.

Methods: The Neurosymbolic Pipeline

The proposed system bridges the gap between raw image data and clinical reasoning through a multi-stage pipeline that prioritizes transparency.

  1. Segmentation & Mesh Generation: The process begins with TotalSegmentator, a deep learning tool used to isolate vertebral bodies. These segments are converted into 3D meshes using the marching cubes algorithm.
  2. Height Map Extraction: To capture the full structural complexity of the bone, the 3D mesh is flattened into a 2D "height map." This map records the height across the entire axial plane, offering a more comprehensive view than traditional point-to-point measurements.
  3. Symbolic Feature Extraction: The height map is partitioned into seven distinct anatomical regions (e.g., anterior, posterior, central). Ratios between these regions are calculated to quantify height loss.
  4. Rule Generation: Using the RuleFit algorithm, the system distills these features into a sparse linear model. This results in a simple, logical formula for diagnosis—for example, flagging a VCF if the anterior height falls below a specific percentage of the posterior height.
The neurosymbolic pipeline for height map extraction and regional segmentation.

Experimental Setup

The study utilized the VerSe19 dataset, a benchmark for vertebrae labeling. After refining the dataset to exclude cases with metal implants or insufficient neighboring vertebrae for comparison, the final cohort consisted of 1,460 annotated thoracolumbar vertebrae from 141 patients.

Key Results and Explainability

The neurosymbolic approach was evaluated against standard deep learning architectures, including ResNeXt50 and DenseNet. The results demonstrate that explainability does not require a performance trade-off.

  • Accuracy: The interpretable model achieved 96% accuracy, outperforming the best-performing DenseNet (95%).
  • Sensitivity: It maintained a high sensitivity of 91%.
  • F1-Score: The model led the field with an F1-score of 0.81, significantly higher than unmasked deep learning baselines (0.62–0.65).

Beyond metrics, the model's primary advantage is its visual decision process. By highlighting exactly which anatomical regions triggered a fracture alert, the system allows radiologists to verify findings in seconds, fostering greater clinical trust.

Visualization of the neurosymbolic decision process and example classification cases.

Discussion and Future Work

This research challenges the myth that AI performance must be sacrificed for interpretability. By utilizing a full 2D height map rather than traditional three-point measurements, the model captures complex "wedge" or "crush" deformations that simpler methods might overlook.

Despite its success, some limitations remain. "Failure" cases often involved labeling noise, where the model identified fractures missed in the ground truth. Future research will focus on clinical integration studies to assess how these explanations affect physician trust and decision-making, as well as broader validation across diverse patient populations and imaging hardware.

Summary updated

BibTeX
@misc{shen2024intrinsically,
  author = {Yiqing Shen and Blanca Inigo Romillo and Yiqing Shen and Benjamin D. Killeen and Michelle Song and Axel Krieger and Christopher Bradley and Mathias Unberath},
  title = {{An Intrinsically Explainable Approach to Detecting Vertebral Compression Fractures in CT Scans via Neurosymbolic Modeling}},
  year = {2024},
  eprint = {2412.17258},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.17258},
}
Figure