A pilot study of deep learning-based CT volumetry for traumatic hemothorax

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Trauma Surgery

Background:

• Traumatic hemothorax (HTX) volume is crucial for predicting hemorrhage-related outcomes.
• Accurate and efficient HTX volume quantification is essential for trauma patient management.
• Current methods rely on subjective expert grading, which can be time-consuming and variable.

Purpose of the Study:

• To employ nnU-Net, a deep learning model, for automated semantic segmentation and quantitative visualization of HTX in trauma patients.
• To assess the performance of automated HTX volumetry using overlap and volume-based metrics.
• To evaluate the accuracy of automated HTX volumes in predicting massive transfusion (MT) and in-hospital mortality (IHM) and compare it to expert consensus.

Main Methods:

• Utilized manually labeled chest CT scans from 77 adult trauma patients with non-negligible HTX.
• Employed ensembled nnU-Net for automated HTX segmentation, validated through fivefold cross-validation.
• Compared automated results with manual segmentations and expert grading using Dice Similarity Coefficient (DSC), volume similarity, Pearson's r, ICC, and AUC for outcome prediction.

Main Results:

• Ensembled nnU-Net achieved high performance with a mean DSC of 0.75 and volume similarity of 0.91.
• The model demonstrated strong correlation with manual measurements (Pearson's r=0.93, ICC=0.92) and minimal overmeasurement bias (1.7 mL).
• Automated HTX volumes showed comparable predictive performance for MT and IHM (AUC=0.74) to manual volumes (AUC=0.76) and expert grading (AUC=0.76).

Conclusions:

• Automated HTX volumetry using nnU-Net demonstrates high method validity and interpretability.
• The AI-driven approach performs comparably to manual quantification and expert consensus in predicting critical hemorrhage outcomes.
• Automated HTX volumetry shows significant promise for advancing research and clinical care in trauma management.