Related Experiment Video
Updated: Aug 22, 2025
03:49
Application of Microwave Ablation in Laparoscopic Partial Splenectomy
Published on: November 15, 2024
262
Toward automated interpretable AAST grading for blunt splenic injury
Haomin Chen1, Mathias Unberath1, David Dreizin2
1Department of Computer Science, Johns Hopkins University, Baltimore, MD, USA.
Emergency Radiology
|November 13, 2022
Summary
This study developed an automated deep learning system to grade splenic injuries from CT scans, achieving substantial agreement with expert consensus. The AI tool rapidly and accurately identifies severe injuries and predicts the need for intervention in trauma patients.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging Analysis
Background:
- The American Association for the Surgery of Trauma (AAST) splenic organ injury scale (OIS) is a standard CT-based grading system for blunt splenic trauma.
- Current inter-rater agreement for AAST grading is modest, highlighting a need for objective, automated diagnostic aids.
- Developing an automated system can improve the consistency and trustworthiness of splenic injury assessments.
Purpose of the Study:
- To pilot the development of an automated, interpretable, multi-stage deep learning system.
- To predict American Association for the Surgery of Trauma (AAST) grades from admission trauma CT scans.
- To create a high-trust diagnostic aid for blunt splenic trauma evaluation.
Main Methods:
- A 4-part pipeline was developed: automated splenic localization, pseudoaneurysm/active bleed detection (Faster R-CNN), splenic parenchymal disruption quantification (nnU-Net), and a directed graph for AAST grade inference.
- The system was trained and validated on 174 adult patients with voxelwise labeling and consensus AAST grading.
- Hemorrhage-related outcome data was used to assess predictive capabilities.
Main Results:
- Substantial agreement (weighted κ = 0.79) was achieved between automated and consensus AAST grades.
- High-grade injuries (IV and V) were predicted with 92% accuracy, 95% positive predictive value, and 89% negative predictive value.
- The automated system demonstrated comparable performance to expert consensus in predicting hemorrhage control intervention (AUC 0.88 vs 0.83) with a mean inference time of 96.9 seconds.
Conclusions:
- The developed automated system is rapid, verifiable, and shows high agreement with expert consensus for AAST grading.
- It accurately diagnoses high-grade splenic lesions and predicts hemorrhage control intervention in adult trauma patients.
- This deep learning approach offers a promising tool for objective and efficient blunt splenic trauma assessment.
