An Interpretable Approach to Automated Severity Scoring in Pelvic Trauma

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Pelvic ring disruptions are common in trauma patients, often requiring classification using the Tile AO/OTA system.
• Automating Tile classification of pelvic fractures from CT scans can enhance radiologist workflow and prioritize urgent cases.

Purpose of the Study:

• To develop an automated and interpretable decision support system for pelvic fracture detection and Tile grade classification.
• To assist trauma radiologists by providing insights from an automated read of whole-body CT scans.

Main Methods:

• A Faster-RCNN model detects pelvic fractures on CT scans with high specificity.
• A structural causal model, based on clinical practices, infers initial Tile grades from detected fractures.
• A Bayesian causal model and object detector refine fracture detection and classification for a final Tile grade.

Main Results:

• The system achieves an AUC of 83.3% for translational instability and 85.1% for rotational instability.
• The automated method provides interpretable features, including fracture location, type, and counterfactuals.
• The approach demonstrates comparable performance to previous black-box methods while enabling human-machine teaming.

Conclusions:

• The developed system offers an automated yet interpretable solution for pelvic fracture grading in trauma CT.
• This AI-driven tool can support radiologists, reduce workload, and improve the efficiency of trauma diagnosis.
• The transparent nature of the system facilitates trust and collaboration between human readers and the AI.