Between July 2018 and October 2020, 2047 routine fetal lung ultrasound images (either right or left lung) were obtained from 2047 women with singleton pregnancies, at gestational ages (GA) ranging from 27+3 up to 42+0 week. All participating women included in the study gave written informed consent for the use of ultrasound images and clinical data. All the methods explained here were performed in accordance with the relevant guidelines and regulations and were approved, together with the study protocol, by the ethics committee of Obstetrics and Gynecology Hospital at Fudan University (2018-73). Of these, 731 babies with GA 28+3– 37+6 weeks were delivered within 72 hours of the ultrasound examination at the hospital. According to the same enrollment criteria of previous studies, the final cohort included 295 women with singleton pregnancies, with a total of 295 fetal-lung ultrasound images. The flow chart for the study population is shown in Fig. 3. Gestational age was determined from the last menstrual period and verified by first-trimester dating ultrasound (crown-to-bottom length).
Pregnancy complications included GDM and PE. GDM was diagnosed using a 75 g oral glucose tolerance test at 24-28 weeks of gestation.27. Pre-eclampsia and gestational hypertension are characterized by new onset hypertension (> 140 mmHg systolic or > 90 mmHg diastolic) after 20 weeks of pregnancy28.
Analysis of neonatal clinical data was supervised by a neonatal physician. NRM included respiratory distress syndrome (RDS) or transient tachypnea of the newborn (TTN). The diagnosis of RDS and TTN is based on symptoms, signs and radiological examination7.29. RDS diagnostic criteria: tachypnea, snoring, chest wall retraction, nasal dilatation, need for supplemental oxygen, and chest X-ray appearance led to admission to the neonatal intensive care unit for respiratory support. Diagnostic criteria of TTN: mild or moderate respiratory distress (isolated tachypnea, infrequent snoring, mild retraction) and a chest radiograph (if performed) showing alveolar and/or pulmonary interstitial effusion and prominent pulmonary vascular patterns.
Ultrasound imaging and segmentation
All ultrasound images were obtained during routine prenatal ultrasound examinations within 72 hours before delivery. Among which, the images of the training group were obtained by radiologist 1 with more than 10 years of experience in obstetric and gynecological ultrasound imaging, using the aWS80A ultrasound system (Samsung, Korea). The frequency of the CA1-7A probe was 1-7 MHz, with a center frequency of 4.0 MHz. Images of the test group were obtained by radiologist 2 with 3 years of experience in obstetric and gynecological ultrasound imaging, using a VOLUSON E8 ultrasound system (GE, United States of America). The frequency of the C1-5-D probe was 2-5 MHz, with a center frequency of 3.5 MHz.
A detailed description of the standard image acquisition protocol and the method used for manual (freehand) determination has been fully described in a previous study.25: Briefly, standard fetal lung imaging requiring: in an axial section of the fetal thorax at the level of the four-chamber cardiac view, settings were adjusted (depth, gain, frequency, and harmony) to ensure that at least one of the lungs was not had visible acoustic shadows from the fetal ribs. All images were inspected for image quality control and saved in DICOM (.dcm) format for offline analysis. Manual delineation (freehand) was performed on each fetal lung by two radiologists (radiologists A and B), and a square outline (40 × 40 pixels) was performed by radiologist B, selecting one side of the fetal lung, taking care great for him. ensure that only lung tissue is outlined and avoiding blood vessels, rib shadows and lung capsule as shown in Fig. 4. Radiologist A’s segmentation results were used to generate the model, while radiologist B’s segmentation and square determination results were used to verify model robustness.
Radiomics Evaluation and Machine Learning
The search process is shown in Figure 5.
All feature extraction and image classifications were performed using Matlab R2018a and the Classification Toolbox (Mathworks, Inc, Natick, Massachusetts, USA).
Univariate analysis was used to describe differences in features between different categories. The t-test was performed on each of the 430 continuous radiomics features25, including 15 morphological, 73 texture, and 342 wavelet features. X2 the test was performed for two categorical clinical characteristics, gestational age and pregnancy complications. P value < 0.05 indicated a significant difference.
The feature extraction method to analyze each ROI has been previously reported25. First, the importance of high-throughput radiomics features for fetal lung imaging was sorted into selected features by varying the out-of-bag data feature of the random regression forest. If a feature is influential, changing its values would affect the error testing of the model with out-of-bin data. The more important a feature is, the greater its impact will be30. As a result, 20 radiomic features (2 texture features and 18 wave features) and 2 clinical features (GA and pregnancy complications) were selected for classification, which are shown in Table 4. Stability of selected radiomic features depending on different definitions (manual determination by radiologists A and B and square determination) was analyzed by ICC (2, 1)31. Next, the diagnostic performance of predicting neonatal respiratory morbidity was compared depending on different characteristics, including clinical characteristics (GA and pregnancy complications), radiomic features, and the combination of clinical and radiomic features. For clinical features, a support vector machine (SVM) classifier was used for classification. By adjusting the cost of misclassification into different categories, the classifier can focus on positive samples. For radiomic features and combination of clinical and radiomic features, with high sample imbalance and small sample size, RUSBoost (Random Subsampling with AdaBoost)32 was used to build the model. Finally, the probability of risk of NRM in each fetal lung image was obtained, which was the predicted score normalized to the interval 0–1 by the softmax function of RUSBoost. The cutoff point of the model was 0.5. Fetal lungs with risk probability higher than 0.5 were divided into the high-risk group and lower than 0.5 were divided into the low-risk group. All classifier parameters were tuned with tenfold bootstrap cross-validation and the decision tree was used as the base learner for RUSBoost.
Model prediction performance was assessed for sensitivity (SENS), specificity (SPEC), accuracy, PPV, NPV and AUC.