Automated fetal magnetic resonance imaging lung segmentation for rapid lung volume estimation across diverse pulmonary hypoplasia phenotypes.

Background

Fetal magnetic resonance imaging (MRI)-derived lung volume measurements are used for prenatal risk stratification in conditions associated with pulmonary hypoplasia, but manual segmentation is time-intensive.

Objective

To evaluate whether automated fetal MRI lung segmentation enables rapid and reliable volumetric assessment across diverse pulmonary hypoplasia phenotypes.

Materials And Methods

In this retrospective study, fetal MRI examinations performed from 2016 to 2025 with expert-adjudicated lung segmentations obtained during clinical care were used for model development and testing. Automated segmentation pipelines were developed for total lung volume (TLV) and percent predicted lung volume (PPLV). Performance was assessed using Dice similarity coefficient (DSC), absolute percentage error (APE), Bland-Altman analysis, intraclass correlation coefficient (ICC), case-level risk-category agreement, and inference time. Lowest-performing TLV outputs were adjudicated by a fetal radiologist to simulate a human-in-the-loop workflow.

Results

The final dataset included 906 fetal MRI examinations with complete and recoverable lung segmentations, acquired at a mean gestational age of 27.1 weeks, range 16-38 weeks. Automated TLV estimation showed excellent agreement with manual volumetry (ICC, 0.955; mean DSC, 0.81±0.10; median APE 11.1%). Automated PPLV estimation showed lower performance (ICC, 0.727; mean DSC, 0.76±0.12; median APE 14.5%). Mean inference time was 4 s per examination. Correction of the 25 lowest-performing TLV outputs required 3.1 min per case.

Conclusion

Automated fetal MRI lung segmentation enables rapid TLV estimation across diverse pulmonary hypoplasia phenotypes and may improve clinical workflow efficiency while preserving expert oversight for challenging cases.