Development of a deep learning-based grading model to evaluate fetal lung maturity from normal fetal lung ultrasound images

Objectives: To develop a deep learning algorithm for quantifying fetal lung maturity from normal fetal lung ultrasound images. Methods This is a single-center analytical study that retrospectively cross-sectionally observes singleton pregnant woman without pregnancy complications at 20–41 + 6 weeks of gestation, and acquires axial images of their fetuses at the level of the four-chambered heart in order to create a model for evaluating the maturity of the fetal lungs using normal fetal lung ultrasound images. Assuming that the texture information of the fetal lung ultrasound image can represent the maturity of the fetal lung development, and the gestational age is proportional to the maturity of the fetal lung, the deep learning grading model based on the normal fetal lung ultrasound image is established with the gestational age as the baseline. The pictures were split into three classes according to the gestational weeks which were calculated at the last menstrual period of the pregnant woman as the reference standard: 20 to 29 + 6 weeks for class I, 30 to 36 + 6 weeks for class II, and 37 to 41 + 6 weeks for class III. Results This study had 350 pregnant women in total, and the deep learning model's classification accuracy for ultrasound images of fetal lungs at different gestational ages was 0.917. Conclusion The grading model based on deep learning can correctly and independently predict the fetal lung maturity from ultrasound images.