An Interpretable Approach to Automated Severity Scoring in Pelvic Trauma

Anna Zapaishchykova, David Dreizin, Zhaoshuo Li, Jie Ying Wu, Shahrooz Faghih Roohi, Mathias Unberath

Introduction

Pelvic ring disruptions represent life-threatening injuries frequently encountered in multi-system trauma, with mortality rates as high as 15%. Clinical management relies heavily on the Tile AO/OTA classification system, which categorizes fractures based on mechanical instability—specifically rotational (R) and translational (T) stability. However, implementing this system in high-pressure trauma environments faces significant hurdles: the sheer volume of whole-body CT scans requiring rapid interpretation, the inherent subjectivity leading to low inter-observer agreement, and the "black box" nature of standard deep learning models which lack the transparency required for high-stakes surgical decisions.

Methods

The proposed solution is a hybrid AI pipeline that separates visual feature extraction from clinical reasoning, mimicking the diagnostic process of a radiologist.

Feature Extraction

The "eyes" of the system consist of a Faster-RCNN architecture trained on 8mm-thick axial Maximum Intensity Projection (MIP) slices. This stage identifies specific anatomical findings, such as pubic symphysis diastasis (PSD) or sacral fractures, which serve as the building blocks for severity grading.

Causal Reasoning and Refinement

The "brain" of the system is a Bayesian structural causal model (BM). This model encodes clinical knowledge regarding how specific fracture types causally relate to the final Tile grade. A novel "refinement loop" allows the system to double-check its work: after identifying high-confidence fractures, the Bayesian model predicts what other injuries should likely be present. The system then re-examines lower-confidence detections from the initial pass to confirm or reject these predicted co-occurring fractures.

Figure 1: Overview of the proposed method, illustrating the two-stage pipeline and the Bayesian refinement loop.

Results

The system was validated using a dataset of 373 admission CT scans from two Level I trauma centers, with ground truth established by a consensus of three trauma-subspecialized radiologists.

The "BM refinement" approach demonstrated superior performance compared to baseline models. The refined model achieved an AUC of 0.85 for rotational instability and 0.83 for translational instability. Furthermore, the system's agreement with experts, measured via Kappa scores, approached the levels of inter-rater variability observed between human experts themselves.

Figure 6: Performance metrics (AUC and Kappa) for rotational (R) and translational (T) instability across different model configurations.

Discussion

A key advantage of this approach is its inherent interpretability. By utilizing a causal model, the system can provide "counterfactual" explanations (e.g., explaining how the grade would change if a specific fracture were present), facilitating better human-machine collaboration. The results challenge the assumption that interpretability requires a sacrifice in accuracy, as the structured model outperformed end-to-end deep learning baselines.

Future work will address the "false-false" trap, where high-confidence false positives from the object detector can bias the final grade, and will involve real-world testing to evaluate how trauma surgeons interact with these interpretable suggestions in clinical practice.

Summary updated

BibTeX
@inproceedings{zapaishchykova2021interpretable,
  author = {Anna Zapaishchykova and David Dreizin and Zhaoshuo Li and Jie Ying Wu and Shahrooz Faghih Roohi and Mathias Unberath},
  title = {{An Interpretable Approach to Automated Severity Scoring in Pelvic Trauma}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  url = {https://openreview.net/forum?id=9MYIco0_Np},
}
Figure