Pulmonary contusion: automated deep learning-based quantitative visualization.

Area of Science:

• Medical imaging and artificial intelligence
• Pulmonary medicine
• Trauma critical care

Background:

• Pulmonary contusion is a significant injury in trauma patients.
• Early prediction of Acute Respiratory Distress Syndrome (ARDS) is crucial for patient management.
• Current methods for quantifying contusion may not be rapid or precise enough.

Purpose of the Study:

• To train and validate deep learning models for automated CT volumetry of pulmonary contusion.
• To calculate the Lung Contusion Index (auto-LCI) as a percentage of total lung volume.
• To assess the association between auto-LCI and clinical outcomes like ARDS, ICU length of stay, and mechanical ventilation duration.

Main Methods:

• Retrospective analysis of 302 adult trauma patients with pulmonary contusion.
• Utilized the nnU-Net deep learning model for automated segmentation of contusions and whole lungs.
• Performed multivariate regression and Cox proportional hazards models incorporating auto-LCI and clinical variables (oxygen saturation, heart rate, blood pressure).

Main Results:

• Deep learning model achieved high accuracy in segmenting pulmonary contusions (Dice score 0.67, ICC 0.90).
• Higher auto-LCI values were significantly associated with ARDS development (p=0.04), longer ICU stays (p=0.02), and increased need for mechanical ventilation (p=0.04).
• The model incorporating auto-LCI and clinical variables showed good predictive performance for ARDS (AUC 0.70).

Conclusions:

• Automated CT volumetry provides a reliable method for quantifying pulmonary contusion.
• Increasing auto-LCI is a significant predictor of ARDS development and resource utilization in trauma patients.
• This AI-driven approach can aid in early risk stratification and clinical decision-making for at-risk trauma patients.