BACKGROUND Pulmonary contusion exists along a spectrum of severity, yet is commonly binarily classified as present or absent. We aimed to develop a deep learning algorithm to automate percent pulmonary contusion computation and exemplify how transfer learning could facilitate large-scale validation. We hypothesized that our deep learning algorithm could automate percent pulmonary contusion computation and that greater percent contusion would be associated with higher odds of adverse inpatient outcomes among patients with rib fractures. METHODS We evaluated admission-day chest computed tomography scans of adults 18 years or older admitted to our institution with multiple rib fractures and pulmonary contusions (2010–2020). We adapted a pretrained convolutional neural network that segments three-dimensional lung volumes and segmented contused lung parenchyma, pulmonary blood vessels, and computed percent pulmonary contusion. Exploratory analysis evaluated associations between percent pulmonary contusion (quartiles) and odds of mechanical ventilation, mortality, and prolonged hospital length of stay using multivariable logistic regression. Sensitivity analysis included pulmonary blood vessel volumes during percent contusion computation. RESULTS A total of 332 patients met inclusion criteria (median, 5 rib fractures), among whom 28% underwent mechanical ventilation and 6% died. The study population's median (interquartile range) percent pulmonary contusion was 4% (2%–8%). Compared to the lowest quartile of percent pulmonary contusion, each increasing quartile was associated with higher adjusted odds of undergoing mechanical ventilation (odds ratio [OR], 1.5; 95% confidence interval [95% CI], 1.1–2.1) and prolonged hospitalization (OR, 1.6; 95% CI, 1.1–2.2), but not with mortality (OR, 1.1; 95% CI, 0.6–2.0). Findings were similar on sensitivity analysis. CONCLUSION We developed a scalable deep learning algorithm to automate percent pulmonary contusion calculating using chest computed tomography scans of adults admitted with rib fractures. Open code sharing and collaborative research are needed to validate our algorithm and exploratory analysis at a large scale. Transfer learning can help harness the full potential of big data and high-performing algorithms to bring precision medicine to the bedside. LEVEL OF EVIDENCE Prognostic and epidemiological, Level III.
No comments:
Post a Comment