The results of the present study revealed that fetal lung texture analysis by ultrasound-based radiomics technology can be used to predict the probability of neonatal respiratory morbidity by analyzing fetal lung ultrasound images and in combination with clinical characteristics (GA and pregnancy complications). It may provide a new method for noninvasive prediction of NRM.
The clinical utility of FLM assays has been largely debated . At present, instead of studying several components of the amniotic fluid through amniocentesis, the application of prenatal corticoids and postnatal surfactant has become the main clinical measure to reduce neonatal respiratory diseases . However, the recommended type of corticosteroid and the gestational window of treatment administration have not been clearly defined . Studies have shown that there are potentially important risks of corticosteroids in neurodevelopment and fetal metabolic planning [14–16]. In a study of 278,508 live-born singletons of 24 weeks gestation or above in Finland, antenatal steroid was shown to be associated with the delivery of small fetus at birth . The results of this study may provide a new method for non-invasive approaches for the prenatal assessment of FLM, which can not only avoid the fear and discomfort of amniocentesis, help to decide whether to use prenatal corticosteroids, but also refine the timing of delivery in high-risk pregnancies.
With the widespread use of ultrasound in obstetrics, several attempts have been made to evaluate fetal lung maturity noninvasively. Kim Sm et al  showed that a measured elevated acceleration-to-ejection time ratio of the fetal pulmonary artery doppler was independently associated with the development of RDS in preterm infants and thus a possible marker of lung maturity. Attempts to quantify fetal lung volume in normal pregnancies by using 3-dimensional ultrasonography though useful in cases like diaphragmatic hernia have not been shown to objectively evaluate FLM [19–20]. In addition, gray scale measurement , fetal lung tissue movement assessment , and evaluation of fetal lung images relative to fetal liver and fetal placenta images  have been tried to proposed as a possible tool for the assessment of fetal lung maturity. Unfortunately, the accuracy of this diagnosis is very poor, so no clinical significance is found. Recently, Palacio M et al.  reported that the quantitative ultrasound lung texture analysis could be used to evaluate fetal lung maturity and showed an accuracy similar to that of biochemical tests in amniotic fluid previously reported. In this study, the overall performance of neonatal respiratory morbidity prediction model based on fetal lung texture analysis by ultrasound-based radiomics technology achieved AUC of 0.81–0.88, sensitivity of 77.78–84.31%, specificity of 81.31–82.09% and accuracy of 81.18–81.90%. These ultrasound images, which appear indistinguishable to the naked eye, could quickly and accurately predict the risk of NRM in the fetus. And the images collected by different trained doctors using different machines do not affect the estimation results of the model.
Our previous research  reported that there were great differences in fetal lung texture between pregnancies with GDM, PE and normal pregnancy and between different gestational ages. In our study population, there were 33.2% (98/295) of pregnant women with GDM and PE. Among these, the proportion of newborns with NRM was nearly twice that in the normal pregnancy group (6.1% vs 3.2%). Therefore, in this study, the model was established by high-throughput radiomics features and two clinical features (pregnancy complications and gestational age). Studies  have shown that the accuracy and PPV of tests on amniotic fluid in predicting NRM was 73.3% (57.5–81.6%) and 27.1% (18.0-34.1%) respectively. In this study, results showed improvements by about 8.2% in accuracy (81.5%) and 29.3% in PPV (56.4%).
Our study had several limitations: First, large amounts of data are necessary in radiomics for mining concealed prognostic information and to avoid overfitting. Expanding the sample size, especially the positive sample size, would improve the stability and accuracy of the model. Second, in this study, the ROIs of fetal lungs were performed manually. A computer system will be used to identify fetal lung tissue automatically, so that the model could be used more conveniently. Third, it is a single-center study, and image acquisition and delineation were performed by highly-trained personnel. But as the number of operators and settings increases, there will be many unqualified images. Multi-center research will be carried out in the future.
In conclusion, ultrasound-based radiomics technology can be used to predict neonatal respiratory morbidity. The results of this study may provide a new method for non-invasive approaches for the prenatal prediction of NRM.