Objective
The objective of the study was to evaluate the performance of automatic quantitative ultrasound analysis (AQUA) texture extractor to predict fetal lung maturity tests in amniotic fluid.
Study Design
Singleton pregnancies (24.0-41.0 weeks) undergoing amniocentesis to assess fetal lung maturity (TDx fetal lung maturity assay [FLM]) were included. A manual-delineated box was placed in the lung area of a 4-chamber view of the fetal thorax. AQUA transformed the information into a set of descriptors. Genetic algorithms extracted the most relevant descriptors and then created and validated a model that could distinguish between mature or immature fetal lungs using TDx-FLM as a reference.
Results
Gestational age at enrollment was (mean [SD]) 32.2 (4.5) weeks. According to the TDx-FLM results, 41 samples were mature and 62 were not. The imaging biomarker based on AQUA presented a sensitivity 95.1%, specificity 85.7%, and an accuracy 90.3% to predict a mature or immature lung.
Conclusion
Fetal lung ultrasound textures extracted by AQUA provided robust features to predict TDx-FLM results.
The most common cause of mortality and neonatal morbidity in preterm and early term fetuses is lung immaturity. The strongest predictor of lung maturity is gestational age, although a quantifiable risk of pulmonary morbidity caused by lung immaturity may be present at any gestational age. Thus, infants who are born at less than 39 weeks have significantly higher rates of neonatal morbidity, including respiratory distress syndrome (RDS) when compared with infants born at a gestation of 39 weeks or longer.
The current methods used to test fetal lung maturity (FLM), including lamellar body count, lecithin-sphingomyelin ratio, or TDx fetal lung maturity assay II test (TDx-FLM II; Abbott Laboratories, Abbott Park, IL), are performed in amniotic fluid and, consequently, require an invasive procedure. Over the last 30 years, the prediction of lung maturity by noninvasive ultrasound methods has been extensively explored. Earlier studies comparing fetal lung echogenicity with the placenta, fetal gut, or liver demonstrated ultrasonographic changes associated with fetal lung maturation. Later studies have used approaches based on quantitative ultrasound analysis to explore the potential of ultrasound to predict fetal lung maturation.
Quantitative ultrasound is based on applying processing methods to ultrasound images. This allows extracting quantitative features and potentially identifying subclinical tissue differences that escape subjective inspection. The use of quantitative ultrasound analysis has previously been investigated for medical diagnostic applications, including breast cancer and liver disease. A few studies have explored the assessment of fetal lung maturity by quantitative ultrasound tissue characterization with different methodologies, with all reporting differences in lung texture features along gestation, individually or compared with the reflection pattern of other organs such as the liver. However, controversial results, limitations in the sample size, and difficulties in recording the parameters proposed with standard means have prevented further development into clinically applicable solutions.
Among various quantitative imaging methods, texture analysis is proposed as a powerful approach to extract quantitative features directly from medical images. We have previously developed an automatic quantitative ultrasound analysis (AQUA) algorithm, which is invariant under illumination changes and does not use direct gray level from the image or tissue references. The method estimates texture features based on conditional random fields so that image texture features converge robustly to identify different tissue characteristics independently of the overall acquisition context, including scanner settings. In a previous study, we demonstrated that AQUA could extract features from fetal lung ultrasound images, showing a strong correlation with gestational age.
In this study, we evaluated the performance of the AQUA texture extractor to predict fetal lung maturity, as assessed by the TDx-FLM II test in amniotic fluid.
Materials and Methods
The study was carried out at the Maternal-Fetal Medicine Department at Hospital Clinic in Barcelona from October 2010 to March 2011. The study population included singleton pregnancies with gestational ages between 24.0 and 41.0 weeks.
Patients were selected from among pregnant women undergoing amniocentesis to assess fetal lung maturity for medical indications or to exclude infection in cases of preterm labor or preterm rupture of membranes. In addition, women scheduled for elective term cesarean section for obstetrical indications were also included. In this latter group, the amniotic fluid was obtained intraoperatively after hysterotomy. Multiple pregnancies and structural/chromosomal anomalies were considered noneligible for this study.
The study protocol was approved by the local Ethics and Institutional Review Board (ID 3823-2007), and all patients provided written informed consent. In all pregnancies gestational age was calculated based on the crown-rump length at first-trimester ultrasound. Maternal and neonatal outcomes were recorded. Descriptive statistics were performed with the SPSS 18.0 statistical software (SPSS Inc, Chicago, IL).
Fetal lung maturity test
A fetal lung maturity was tested in amniotic fluid using the surfactant-to-albumin ratio by fluorescence polarization (TDx-FLM II). Gestational age–specific cutoff values were used to classify the results as mature or immature. This adjustment for gestational age results in a significant improvement in the capacity of the test to predict the respiratory distress syndrome and may simplify clinical decisions.
Image acquisition and delineation of lung tissue
Ultrasound images were obtained the same day of amniotic fluid collection in all cases. A semilateral transverse 4 cardiac-chamber view plane for lung image acquisition was performed using Siemens Sonoline Antares (Siemens Medical Systems, Malvern, PA) and Voluson 730/780 Pro (GE Medical Systems, Milwaukee, WI) ultrasound equipment, all equipped with a curved linear transducer with a frequency range from 3 to 7.5 MHz. Equipment settings were adjusted at the discretion of the clinician performing the ultrasound to obtain the optimal image quality according to clinical criteria.
Images were digitally collected in the original Digital Imaging and Communication in Medicine format and stored for off-line analysis. Manual delineation was performed in the proximal lung (according to the distance to the ultrasound transducer) with a custom-made program with a Graphic User Interface tool developed with MATLAB R2007b (version 7.5.0.342; MATLAB; The MathWorks Inc, Natick, MA). Care was taken to delineate only lung tissue as illustrated in Figure 1 .
AQUA and machine learning prediction
AQUA software was applied to extract the texture features from the manually delineated lungs. Once AQUA was executed, each image became a set of descriptors (15,300 per delineated lung), which include a high amount of textural-related information per delineation. Dimensionality reduction and machine learning-based models were then applied to select a set of descriptors that correlated robustly with the results of the FLM test in amniotic fluid. For this purpose, genetic algorithms and support vector machines were applied.
Genetic algorithms are computational techniques used to resolve the problem of time that would need an exhaustive algorithm to select every possible combination of the 15,300 descriptors to create and validate a predictor. Support vector machines (SVMs) are supervised machine learning algorithms that can be used to predict or classify sets of labeled objects. These SVMs work in 2 phases. Phase 1 is called training, and at this stage, a subset of objects (represented by their feature vectors and the class or label they belong to) are given as inputs to the system. Using the feature vectors, the SVM builds a mathematical model. In phase 2, which is called testing or validation, the model automatically determines the class or label of the new objects. The success rate in this stage is measured, and therefore, the accuracy of the model created is expressed.
In the present study, genetic algorithms identified a set of 31 critical descriptors, which were used in subsequent steps. The support vector machines were then used to develop a mathematical model that combined the set of 31 descriptors to evaluate their theoretical predictive value of the results of lung maturity in amniotic fluid. The association and predictive value were tested with a strategy consisting in the definition of 2 randomly created datasets, containing the same number of women, “A” and “B” with a similar number of mature and immature fetuses. The set of 31 descriptors was used to train a machine learning model.
The accuracy of the model proposed was double tested: the first model was trained with data set B and validated in A. Subsequently, the groups were interchanged and a second model was trained with dataset A and validated in B using the same 31 descriptors. The mean accuracy resulting from the 2 validation tests was calculated and stored as the fitness.
Results
A total of 103 women from 24.0 to 41.0 weeks of gestation were included in the study. The baseline characteristics of the study population are described in the Table . Considering gestational age cutoff values of the TDx-FLM test, 62 and 41 cases were classified as immature or mature, respectively.
| Variable | Mean (SD) or n (%) |
|---|---|
| Maternal age, y | 31.4 (5.6) |
| Nuliparity | 60 (58.3) |
| Gestational age at amniocentesis, wks | 32.2 (4.5) |
| 24 to <28 | 19/103 (18.5) |
| 28 to <32 | 26/103 (25.2) |
| 32 to <34 | 21/103 (20.4) |
| 34 to <37 | 14/103 (13.6) |
| ≥37 | 23/103 (22.3) |
| Indication for amniocentesis | |
| Preterm labor | 41/103 (39.8) |
| Preeclampsia/IUGR | 8/103 (7.8) |
| PPROM | 29/103 (28.2) |
| Methrorrhagia | 4/103 (3.9) |
| Elective | 21/103 (20.4) |
| Gestational age at delivery, wks | 34.3 (4.1) |
| Delivery <37 | 63/101 (62.4) |
| Delivery <34 | 43/101 (42.6) |
| Delivery <28 | 6/101 (5.9) |
| Birthweight, g | 2322 (976) |
| pH AU <7.10 | 3/73 (4.1) a |
| Apgar 5 minutes <7 | 4/101 (4.0) |
| NICU admission | 56/101 (54.5) |
| Respiratory distress syndrome | 17/101 (16.8) |
| Days at NICU, d | 19.8 (21.0) b |
| Neonatal death | 5/101 (5.0) |
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
