Purpose or Case Report: To develop an optimized AI model to automatically segment lung volumes from pulmonary magnetic resonance images (MRI) and generate tidal volume calculations for neonatal patients with chronic lung disease of prematurity (bronchopulmonary dysplasia, BPD).
Methods & Materials: 145 ungated and 27 gated ultrashort echo-time MRI examinations from 133 neonatal patients (13 controls, 14 with non-BPD related indications, and 105 with varying severities of BPD) were used for training . 20% of the training data was reserved for validation during training. Lung volumes were manually segmented from these examinations. Multiple state-of-the-art deep learning architectures for medical image segmentation available through the MONAI (Medical Open Network for AI) framework were experimented. Weights and Biases experiment tracking system along with its hyper-parameter search functionality was used to determine the optimal initial learning rates, learning rate schedulers, optimizers, and data augmentation strategies. The model with the best dice score during training was tested on an independent test dataset consisting of 31 gated MRI examinations from patients with severe BPD, where the respiratory cycle was represented by 2 frames (end of expiration and end of inspiration). The tidal volumes were calculated by subtracting lung volumes of the frames representing end of inspiration from those representing end of expiration. Ground truth lung segmentations of the end of expiration frames of these gated exams were manually generated using 3D Slicer and used for model testing. Relative volume error and Dice coefficients were calculated to evaluate model performance.
Results: MONAI’s dynamic U-net architecture with a linear learning rate scheduler and initial learning rate of 0.0002 with weighted Adam optimizer was identified as the model with the best validation metrics. The mean relative volume error (± standard deviation) of the model on the independent test set was 4.2% (± 3.3%). The mean (± standard deviation) Dice coefficient was 0.92 (± 0.05). The mean (± standard deviation) tidal volume was 12 (± 5.9) ml.
Conclusions: This study reports the tidal volume, an important clinical parameter of lung function, in a cohort of neonates with BPD. The automatic segmentation of lung volume can aid in the calculation of useful clinical parameters, including tidal volume, that characterize lung function in neonates. This could play an important role in clinical decision making for BPD diagnosis and treatment.