An Interpretable Approach to Automated Severity Scoring in Pelvic Trauma

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

TL;DR

Researchers have developed a hybrid AI system that combines deep learning with causal Bayesian networks to grade pelvic fracture severity from CT scans. By mimicking the reasoning process of radiologists, the system provides transparent, interpretable results that match the performance of "black-box" models while enabling better human-machine collaboration in high-stakes trauma settings.

What Problem Does This Paper Address?

Pelvic ring disruptions are life-threatening injuries often found in multi-system trauma patients, with mortality rates reaching up to 15%. To manage these patients, radiologists use the Tile AO/OTA classification system, which grades fractures based on mechanical instability (rotational vs. translational).

However, there are three major hurdles:

  1. Volume and Speed: Busy trauma centers generate a massive volume of whole-body CTs that need immediate interpretation.
  2. Subjectivity: Grading is complex; even experienced radiologists often disagree on the exact Tile grade (low inter-observer agreement).
  3. The "Black Box" Problem: While deep learning can automate this, most models are opaque. In a trauma bay, a surgeon needs to know why a system recommends a specific grade before making a high-stakes surgical decision.

Methods

The authors propose a two-stage pipeline that separates "seeing" from "reasoning":

  1. Feature Extraction (The "Eyes"): The system uses a Faster-RCNN (a standard object detection architecture) trained on axial Maximum Intensity Projection (MIP) slices. It looks for specific findings like pubic symphysis diastasis (PSD) or sacral fractures. To simplify the 3D CT data without losing too much information, they used 8mm-thick MIP slices.
  2. Causal Reasoning (The "Brain"): Instead of a standard classifier, they use a Bayesian structural causal model (BM). This model is built on clinical best practices, encoding how different fracture types (e.g., an ischial spine avulsion) causally relate to the final Tile grade.
  3. The Refinement Loop: This is the novel part. The system first identifies high-confidence fractures. It then uses the Bayesian model to ask: "If these fractures exist, what else should I likely see?" It then goes back to the "lower-confidence" detections from the Faster-RCNN to see if those predicted co-occurring fractures are present, effectively "double-checking" its own work.

Datasets

The study used a dataset of 373 admission CT scans of adult patients from two Level I trauma centers.

  • Ground Truth: Annotations were performed by three trauma-subspecialized radiologists, with a consensus vote used as the gold standard.
  • Features: The dataset included labels for specific fracture types (PSD, SI joint diastasis, etc.) and patient-level Tile grades.

Key Insights and Ideas

  • Mimicking Human Expertise: The refinement loop mimics how a radiologist works—if they see a specific posterior injury, they know to look more closely for a subtle anterior ring fracture that might have been missed at first glance.
  • Interpretability via Counterfactuals: Because the system uses a causal model, it can provide "counterfactuals." It can tell a doctor: "I graded this as Tile B, but if you find a sacral fracture here, it would become a Tile C." This allows for much more effective human-machine teaming.
  • No Performance Trade-off: A common myth in AI is that you must sacrifice accuracy for interpretability. This paper demonstrates that a structured, transparent model can perform just as well as (or better than) an end-to-end deep learning model.

Key Results

The system was evaluated on its ability to detect translational (T) and rotational (R) instability:

  • Performance: The refined model achieved an AUC of 85.1% for rotational instability and 83.3% for translational instability.
  • Reliability: The "BM refinement" approach significantly outperformed models that simply used all detections (high and low confidence) without the causal logic.
  • Agreement: The system's agreement with experts (measured via Kappa scores) reached levels close to the inter-rater variability between human experts themselves.

Limitations and Future Work

  • The "False-False" Trap: The authors note a limitation: if the object detector identifies a "false positive" with very high confidence, the Bayesian model currently has no mechanism to reject it. This could irrevocably bias the final grade.
  • Human-in-the-loop Testing: While the system is designed for human-machine teaming, this specific paper focuses on the algorithmic performance. Future work is needed to see how attending radiologists actually interact with these "interpretable" suggestions in a real-world, high-pressure environment.
  • Expansion: The core idea—combining deep learning "features" with a causal "reasoner"—could likely be applied to many other complex diagnostic tasks beyond pelvic trauma.

Summary updated

Abstract

Pelvic ring disruptions result from blunt injury mechanisms and are often found in patients with multi-system trauma. To grade pelvic fracture severity in trauma victims based on whole-body CT, the Tile AO/OTA classification is frequently used. Due to the high volume of whole-body trauma CTs generated in busy trauma centers, an automated approach to Tile classification would provide substantial value, e. g., to prioritize the reading queue of the attending trauma radiologist. In such scenario, an automated method should perform grading based on a transparent process and based on interpretable features to enable interaction with human readers and lower their workload by offering insights from a first automated read of the scan. This paper introduces an automated yet interpretable pelvic trauma decision support system to assist radiologists in fracture detection and Tile grade classification. The method operates similarly to human interpretation of CT scans and first detects distinct pelvic fractures on CT with high specificity using a Faster-RCNN model that are then interpreted using a structural causal model based on clinical best practices to infer an initial Tile grade. The Bayesian causal model and finally, the object detector are then queried for likely co-occurring fractures that may have been rejected initially due to the highly specific operating point of the detector, resulting in an updated list of detected fractures and corresponding final Tile grade. Our method is transparent in that it provides finding location and type using the object detector, as well as information on important counterfactuals that would invalidate the system’s recommendation and achieves an AUC of 83.3%/85.1% for translational/rotational instability. Despite being designed for human-machine teaming, our approach does not compromise on performance compared to previous black-box approaches.
BibTeX
@inproceedings{zapaishchykova2021interpretable,
  author = {Anna Zapaishchykova and David Dreizin and Zhaoshuo Li and Jie Ying Wu and Shahrooz Faghihroohi 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