2023 ARRS ANNUAL MEETING - ABSTRACTS

RETURN TO ABSTRACT LISTING


E2350. Pulmonary Contusion: Automated Quantitative Visualization with Deep Learning
Authors
  1. Nathan Sarkar; University of Maryland
  2. Lei Zhang; University of Maryland
  3. Yuanyuan Liang; University of Maryland
  4. Peter Campbell; University of Maryland
  5. Udit Khetan; University of Maryland
  6. Mustafa Khedr; University of Maryland
  7. David Dreizin; University of Maryland
Objective:
Pulmonary contusions are a common computed tomography (CT) finding in thoracic trauma, and previous reports have shown that the extent of pulmonary contusion, measured semi-automatically, is a risk factor for the development of Acute Respiratory Distress Syndrome (ARDS). Rapid fully-automated quantitative visualization of pulmonary contusion is needed for point of care use. This study aims to 1) train and validate a deep learning (DL) model for segmenting pulmonary contusion and 2) assess the relationship between automated contusion volumes and ARDS.

Materials and Methods:
Patients with pulmonary contusion (n = 302) were identified from reports between 2016 and 2021. The pulmonary contusion volume was then manually segmented. This segmentation data was used to train nnU-net, a state-of-the-art ensemble DL method in five-fold cross-validation. Clinical data was obtained from the medical record. Criteria for exclusion from the clinical analysis included absence of records sufficient to determine ARDS status, and catastrophic head injury (Abbreviated Injury Scale of 6). DL model performance was evaluated with a combination of overlap and volume-based metrics including Dice similarity coefficient (DSC), volume similarity index (VSI), Pearson’s r, and interclass correlation coefficient (ICC). Volumes were divided into quartile ranges. Logistic regression was used to determine ARDS risk, and a Cox proportional hazards model was used to assess differences in mechanical ventilation time at a 95% confidence level.

Results:
Mean ground-truth volume was 757 mL (range: 4 - 6633 mL) and mean automated volume was 706 mL (range: 0 - 7181 mL). Volume Similarity Index and mean Dice score were 0.82 and 0.67. Interclass correlation coefficient and Pearson r (manual versus automated) were 0.90 and 0.91. Of the 280 patients included in the clinical analysis, 38 patients developed ARDS. Automated contusion volume was associated with ARDS (p = 0.003) and need for mechanical ventilation (p < 0.0001). At volumes of 100, 1000, and 3000 mL, % probability of ARDS was 8%, 17%, and 35% respectively. Those in the top quartile (> 890 mL) had 4.2-fold higher odds of developing ARDS, and 4.4-fold higher odds of requiring mechanical ventilation compared to the bottom quartile (< 95 mL) and required significantly longer ventilator times in the subset of mechanically vented patients (p = 0.002). Factors also contributing to ARDS development included admission systolic blood pressure (p = 0.02), admission O2 saturation (p = 0.01), and the injury severity score (ISS) (p < 0.001).

Conclusion:
Automated DL-based quantitative visualization of pulmonary contusion yielded reliable results. Patients with higher volumes were at greater risk of ARDS and require longer periods of mechanical ventilation.