Automated MRI Lung Segmentation and 3D Morphologic Features for Quantification of Neonatal Lung Disease

Abstract

Purpose: To analyze the performance of deep learning (DL) models for segmentation of the neonatal lung in MRI and investigate the use of automated MRI-based features for assessment of neonatal lung disease.

Materials and methods: Quiet-breathing MRI was prospectively performed in two independent cohorts of preterm infants (median gestational age, 26.57 weeks; IQR, 25.3-28.6 weeks; 55 female and 48 male infants) with (n = 86) and without (n = 21) chronic lung disease (bronchopulmonary dysplasia [BPD]). Convolutional neural networks were developed for lung segmentation, and a three-dimensional reconstruction was used to calculate MRI features for lung volume, shape, pixel intensity, and surface. These features were explored as indicators of BPD and disease-associated lung structural remodeling through correlation with lung injury scores and multinomial models for BPD severity stratification.

Results: The lung segmentation model reached a volumetric Dice coefficient of 0.908 in cross-validation and 0.880 on the independent test dataset, matching expert-level performance across disease grades. MRI lung features demonstrated significant correlations with lung injury scores and added structural information for the separation of neonates with BPD (BPD vs no BPD: average area under the receiver operating characteristic curve [AUC], 0.92 ± 0.02 [SD]; no or mild BPD vs moderate or severe BPD: average AUC, 0.84 ± 0.03).

Conclusion: This study demonstrated high performance of DL models for MRI neonatal lung segmentation and showed the potential of automated MRI features for diagnostic assessment of neonatal lung disease while avoiding radiation exposure.Keywords: Bronchopulmonary Dysplasia, Chronic Lung Disease, Preterm Infant, Lung Segmentation, Lung MRI, BPD Severity Assessment, Deep Learning, Lung Imaging Biomarkers, Lung Topology Supplemental material is available for this article. Published under a CC BY 4.0 license.See also the commentary by Parraga and Sharma in this issue.

Keywords: BPD Severity Assessment; Bronchopulmonary Dysplasia; Chronic Lung Disease; Deep Learning; Lung Imaging Biomarkers; Lung MRI; Lung Segmentation; Lung Topology; Preterm Infant.

Graphical abstract

Figure 1:

Data flow and participant exclusion process for the analyses performed in this study. n = number of participants analyzed. APGAR = appearance, pulse, grimace, activity, and respiration, BPD = bronchopulmonary dysplasia, DL = deep learning, FRC = functional residual capacity, ILFT = infant lung function testing.

Figure 2:

MRI-based neonatal lung segmentation and automated MRI analysis. (A) Clinical study including preterm infants with and without bronchopulmonary dysplasia (BPD). Free-breathing neonatal MRI was performed at mean gestational age of 37 weeks ± 5.8. (B) Manual MRI annotation of the lung was performed by three trained physicians (physician 1 [P1], physician 2 [P2], and physician 3 [P3]). MRI morphologic injuries (eg, emphysema, fibrosis, ventilation inhomogeneity) were scored by two trained physicians. (C, D) U-Net deep learning models (MP1, MP2, MP3) were trained for lung segmentation, and a final lung-mask prediction was calculated with an ensemble of the models (ME) through majority voting. (E) Lung volume three-dimensional (3D) reconstruction and automated calculation of 78 lung morphologic 3D descriptors.

Figure 3:

Lung segmentation and lung volume analysis. (A) MRI lung segmentation sample with manual annotation (magenta) and machine learning model–generated lung masks (cyan). (B) Plot shows lung segmentation performances for manual physician-based lung annotations (physician 1 [P1], physician 2 [P2], physician 3 [P3]), and the model ensemble (ME) with majority voting; results are separated for cohort 1 and cohort 2. Boxes represent IQR (25th–75th percentile), median value is the horizontal midline, whiskers extend to data points within ± 1.5 IQR from each quartile, outliers are plotted as diamonds. (C) Graph shows MRI lung volume from the U-Net model ensemble segmentations versus estimated lung volume from manual segmentations (n = 107). (D) Graph shows functional residual capacity per birth weight versus MRI model ensemble lung volume per birth weight (n = 27). (E) Graph shows tidal volume per birth weight versus MRI model ensemble lung volume per birth weight (n = 32). The shaded area in C–E corresponds to the regression 95% CI, and axis plots show univariate histograms and probability density curves.

Figure 4:

Correlation of MRI lung features with bronchopulmonary dysplasia (BPD) severity and lung injury scores. (A) Plot shows predicted lung volume normalized by birth weight against BPD severity grades (n = 103). (B) Graph shows correlation of lung volume based on model ensemble segmentations normalized by birth weight against duration of mechanical ventilation (in days) (n = 103). (C) Graph shows correlation of lung volume based on model ensemble segmentations normalized by birth weight with duration of oxygen supplementation (in days) (n = 103). (D) Plot shows lung elongation (major axis/minor axis) by BPD severity for right and left lungs (n = 103). The shaded area in B and C corresponds to the regression 95% CI. (E) Plot shows MRI lung intensity anteroposterior (AP) centroid displacement versus anteroposterior gradient score for ventilation inhomogeneity (n = 58). (F) Plot shows MRI lung volumetric surface roughness versus interstitial lung injury score for fibrosis (n = 58). max = maximum. Differences were tested with the Kruskal-Wallis H test with Bonferroni multiple test correction, and pairwise comparisons were performed with the Mann-Whitney U test (* = P ≤ .05, ** = P ≤ .01, *** = P ≤ .001). Boxes in A and D–F represent IQR (25th–75th percentile), median value is the horizontal midline, whiskers extend to data points within ± 1.5 IQR from each quartile, outliers are plotted as diamonds.

Figure 5:

Correlation matrix of three-dimensional (3D) MRI lung features with clinical variables (bronchopulmonary dysplasia [BPD] diagnosis variables and BPD risk factors) and lung injury scores. MRI lung features are grouped by feature type (volumetric, intensity, and surface). A subset of morphologic features with the highest Spearman correlations is shown. Statistical significance is annotated based on Spearman correlations with multiple test Bonferroni correction (* = P ≤ .05, ** = P ≤ .01, *** = P ≤ .001). AP = anteroposterior, CC = craniocaudal, gest = gestational, resp = respiratory.

Figure 6:

Bronchopulmonary dysplasia (BPD) classification with best performing models by feature group (GA = gestational age, L = 78 MRI automated lung features, PC = patient and clinical variables, PCL = patient, clinical, and lung features). (A) Plot shows BPD binomial classification performance (no or mild vs moderate or severe). (B) Plot shows BPD multinomial classification performance (no, mild, moderate, severe). (C) Graph shows BPD multinomial receiver operating characteristic (ROC) curve for the best model with PCL features. (D) Plot shows regression performance for duration of respiratory support. (E) Plot shows BPD multinomial classification performance for patients with GA between 25.4 and 28.6 weeks. (F) Graph shows BPD multinomial receiver operating characteristic curve (ROC) for the best model with PCL features with GA (25.4–28.6 weeks). AUC = area under the receiver operating characteristic curve, Log. Reg. = logistic regression, PCA = principal component analysis, RF = random forest, UFS = univariate feature selection. Differences were tested with the Kruskal-Wallis H test with Bonferroni multiple test correction, and pairwise comparisons were performed with the Mann-Whitney U test (* = P ≤ .05, ** = P ≤ .01, *** = P ≤ .001). Boxes in A, B, D, and E represent IQR (25th–75th percentile), median value is the horizontal midline, whiskers extend to data points within ± 1.5 IQR from each quartile, and outliers are plotted as diamonds.