News Science & Technology

Ultrasound-based radiomics technology in fetal lung texture analysis prediction of neonatal respiratory morbidity


Between July 2018 and October 2020, 2047 routine fetal-lung ultrasound images (either right or left lung) from 2047 women with singleton pregnancy were obtained, at gestational ages (GA) ranging from 27+3 to 42+0 weeks. All participating women included in the study gave written informed consent for the use of ultrasound images and clinical data. All the methods hereby explained were performed in accordance with the relevant guidelines and regulations and approved, together with the study protocol, by the ethics committee of the Obstetrics and Gynecology Hospital Affiliated to Fudan University (2018-73). Of these, 731 babies with GA 28+3–37+6 weeks were delivered within 72 h after ultrasound examination in the hospital. According to the same enrolment criteria of previous studies, the final cohort comprised 295 women with singleton pregnancy, with a total of 295 fetal-lung ultrasound images. The flowchart for the study population is shown in Fig. 3. Gestational age was determined by last menstrual period and verified by first-trimester dating ultrasound (crown–rump length).

Figure 3
figure 3

Flowchart of the selection of the study population. NRM neonatal respiratory morbidity.

Pregnancy complications included GDM and PE. GDM was diagnosed using a 75-g oral glucose tolerance test at 24–28 weeks of gestation27. Pre-eclampsia and gestational hypertension are characterized by the new onset of hypertension (> 140 mmHg systolic or > 90 mmHg diastolic) after 20 weeks gestation28.

Analysis of neonatal clinical data was supervised by a neonatal doctor. 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. Diagnostic criteria of RDS: tachypnea, snoring, chest wall retraction, nasal dilatation, the need for supplemental oxygen and the appearance of chest X-rays led to admission to the neonatal intensive care unit for respiratory support. Diagnostic criteria of TTN: mild or moderate respiratory distress (isolated tachypnea, rare snoring, slight retraction) and a chest X-ray (if done) 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 h before delivery. Among which, the images of the training set were obtained by radiologist 1 with more than 10 years’ experience in obstetric and gynecological ultrasound imaging, using aWS80A ultrasound system (Samsung, Korea). The frequency of the CA1-7A probe was 1–7 MHz, with a center frequency was 4.0 MHz. The images of the testing set were obtained by radiologist 2 with 3 years’ experience in obstetric and gynecological ultrasound imaging, using a VOLUSON E8 ultrasound system (GE, United States) . The frequency of the C1-5-D probe was 2–5 MHz, with a center frequency was 3.5 MHz.

A detailed description of the standard image acquisition protocol and the method used of manual (free-hand) delineation is fully described in a previous study25: Briefly, the standard fetal lung images requiring: on an axial section of the fetal thorax at the level of the four-chamber cardiac view, the settings were adjusted (depth, gain, frequency and harmonics) to ensure that at least one of the lungs had no obvious acoustic shadowing from the fetal ribs. All the images were inspected for image quality control and stored in DICOM format (.dcm) for offline analysis. Manual (free-hand) delineation was performed in each fetal lung by two radiologists (radiologists A and B), and square delineation (40 × 40 pixels) was performed by radiologist B, selecting one side of the fetal lung, taking great care to ensure that only the lung tissue was delineated, and avoiding blood vessels, rib shadows, and the lung capsule, as shown in Fig. 4. The radiologist A’s segmentation results were used to generate the model, while the radiologist B’s segmentation and the square delineation results were utilized to verify the stability of the model.

Figure 4
figure 4

Fetal human lung ultrasound images with defined regions of interest. (a,a1,a2,a3) Are images of training set. (b,b1,b2,b3) Are images of testing set. (a1,b1) Manual delineation (radiologist A) of each lung. (a2,b2) Manual delineation (radiologist B) of each lung. (a3,b3) Square delineation (40 × 40 pixels) of each lung. (a,a1,a2,a3) Image of left lung at 36+1 weeks in woman with pre-eclampsia (PE). Cesarean delivery occurred 2 days after ultrasound examination, and baby was diagnosed with transient tachypnea of the newborn. The risk probability derived from the model is 0.829 (> 0.5). (b,b1,b2,b3) Image of left lung at 34+0 weeks in woman with gestational diabetes mellitus (GDM). Cesarean delivery occurred immediately after ultrasound examination, and baby was diagnosed with respiratory distress syndrome. The risk probability derived from the model is 0.843 (> 0.5).

Radiomics evaluation and machine learning

The research process is shown in Fig. 5.

Figure 5
figure 5

Workflow of the fetal lung texture analysis system based on ultrasound-based radiomics technology. Stage I: Fetal-lung US image (four-chamber view) was segmented manually. Stage II: 430 high-throughput radiomics features were extracted from each segmented image. Then features were selected by permuting out-of-bag data feature of random regression forest. And the prediction model was built using RUSBoost (Random under-sampling with AdaBoost). Finally, the risk probability of NRM in each fetal lung image was obtained and divided into the high-risk group or low-risk group. Stage III: According to results of confusion matrix, performance of prediction model was assessed by sensitivity (SENS), specificity (SPEC), accuracy (ACC) and area under receiver-operating-characteristics (ROC) curve. ROI Region of interest, US ultrasound, NRM neonatal respiratory morbidity, Sens sensitivity, Spec specificity, Acc accuracy, ROC receiver operating characteristics.

All the feature extraction and image classifications were carried out using Matlab R2018a and Toolbox Classification (Mathworks, Inc, Natick, Massachusetts, US).

Univariate analysis was used to describe the differences in features among the different categories. The t-test was performed on each 430 continuous radiomics features25, including 15 morphological, 73 texture and 342 wavelet features. The χ2 test was performed for two categorical clinical features, 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, high-throughput radiomics features importance per fetal lung image were ranked to selected features by permuting out-of-bag data feature of random regression forest. If a feature is influential, permuting its values would influence the model error testing with out-of-bag data. The more important a feature is, the greater its influence will be30. As a result, 20 radiomics features (2 texture features and 18 wavelet features) and 2 clinical features (GA and Pregnancy complications) were selected to classification, which are shown in Table 4. The stability of selected radiomic features depending on different delineations (manual delineation by radiologists A and B and square delineation) was analyzed with ICC (2, 1)31. Then, the diagnostic performance of predicting neonatal respiratory morbidity depending on different features was compared, including clinical features (GA and pregnancy complications), radiomics features and the combination of clinical and radiomics features. For clinical features, a support vector machine (SVM) classifier was used for classification. By adjusting the cost of misclassification in different categories, the classifier can focus on the positive samples. For radiomics features and the combination of clinical and radiomics features, with the high imbalance of samples and the small sample size, RUSBoost (Random under-sampling with AdaBoost)32 was used to build the model. Finally, the risk probability of NRM in each fetal lung image was obtained, which was the predicted score normalized to the range of 0–1 by softmax function of the RUSBoost. The cut-off point of the model was 0.5. The 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 bootstrap tenfold cross-validation, and the decision tree was employed as the base learner for RUSBoost.

Table 4 List of high-throughput sonographic features.

The prediction performance of the model was assessed for sensitivity (SENS), specificity (SPEC), accuracy, PPV, NPV and AUC.

Source link