Deep Learning-Based Magnetic Resonance Imaging Lung Segmentation and Volumetric Marker Extraction in Preterm Infants

The diagnosis of neonatal respiratory diseases is currently based on clinical criteria. However, lung structural information is generally lacking due to the unavailability of routinely applicable, radiation-free imaging tools as well as the time-consuming, often non-standardized manual analysis of imaging data. Increased efficiency, comparability and accuracy in image quantification is needed in this patient cohort as pulmonary complications determine immediate and long-term survival. We therefore developed an ensemble of deep convolutional neural networks to perform lung segmentation in magnetic resonance imaging (MRI) sequences obtained in premature infants near term (n=107), with subsequent reconstruction of the 3-dimensional neonatal lung and estimation of MRI lung descriptors for volume, shape, surface, and signal intensity distribution. Annotation of lung segments in quiet-breathing MRI for infants with and without Bronchopulmonary Dysplasia (BPD) was achieved by development of a deep learning model reaching a volumetric dice score (VDC) of 0.908 and validated in an independent cohort (VDC 0.880), thereby matching expert-level performance while demonstrating transferability, robustness towards technical (low spatial resolution, movement artifacts) and lung disease grades. MRI lung descriptors presented relevant correlations with lung lesion scores and enabled the separation of neonates with and without BPD (AUC 0.92{+/-}0.016), mild vs severe BPD (AUC 0.84{+/-}0.027), and single level prediction of BPD severity (AUC 0.75{+/-}0.013). Our work demonstrates the potential of AI-supported MRI markers as a diagnostic tool, characterizing changes in lung structure in neonatal respiratory disease while avoiding radiation exposure.