Automatic Lung Aeration Assessment for Lung Ultrasound Imaging in the Pediatric Intensive Care Unit

Imaging of the lungs is traditionally based on chest X-ray as a first-line imaging method for lung aeration assessment. However, radiation exposure limits its use for patients in the pediatric intensive care unit. Lung ultrasonography (LUS) is a suitable alternative, but its interpretation is highly observer-dependent and requires sufficient experience and skill. We sought to develop a model based on deep learning to assist the clinician in the interpretation of LUS observations. In this retrospective, single-center, proof-of-concept study, all patients, age 0–18 years old admitted at the PICU of the Leiden University Medical Center (LUMC) between January and May 2022 who underwent an LUS were included. LUS video frames were analyzed using a deep learning tool; a conditional generative adversarial network (cGAN) was developed to generate segmentation masks containing clinical features from individual LUS frames. A total of 31 patients, with a median age of 2.5 months (IQR 0–11 months), were analyzed. A total of 98 LUS assessments and 506 4-s videos were collected. The median LUS score was 12 (IQR 8–17). The two best-performing frame-based segmentation models achieved mean Dice similarity coefficients of 0.97 ± 0.03 and 0.96 ± 0.03, with mean squared errors of 0.025 ± 0.025 and 0.030 ± 0.026, respectively. These findings demonstrate that a pediatric-specific cGAN can segment key LUS features from individual frames. However, the results derive from a small, single-center cohort with a frame-level rather than patient-level data split, and no formal clinical validation; independent, prospectively collected validation cohorts are required before any clinical implementation.