Three-dimensional volumetry and fetal weight measurement

Deep learning-based segmentation of whole-body fetal MRI and fetal weight estimation: assessing performance, repeatability, and reproducibility

OBJECTIVES

To develop a deep-learning method for whole-body fetal segmentation based on MRI; to assess the method's repeatability, reproducibility, and accuracy; to create an MRI-based normal fetal weight growth chart; and to assess the sensitivity to detect fetuses with growth restriction (FGR).

METHODS

Retrospective data of 348 fetuses with gestational age (GA) of 19-39 weeks were included: 249 normal appropriate for GA (AGA), 19 FGR, and 80 Other (having various imaging abnormalities). A fetal whole-body segmentation model with a quality estimation module was developed and evaluated in 169 cases. The method was evaluated for its repeatability (repeated scans within the same scanner, n = 22), reproducibility (different scanners, n = 6), and accuracy (compared with birth weight, n = 7). A normal MRI-based growth chart was derived.

RESULTS

The method achieved a Dice = 0.973, absolute volume difference ratio (VDR) = 1.8% and VDR mean difference = 0.75% ([Formula: see text]: - 3.95%, 5.46), and high agreement with the gold standard. The method achieved a repeatability coefficient = 4.01%, ICC = 0.99, high reproducibility with a mean difference = 2.21% ([Formula: see text]: - 1.92%, 6.35%), and high accuracy with a mean difference between estimated fetal weight (EFW) and birth weight of - 0.39% ([Formula: see text]: - 8.23%, 7.45%). A normal growth chart (n = 246) was consistent with four ultrasound charts. EFW based on MRI correctly predicted birth-weight percentiles for all 18 fetuses ≤ 10thpercentile and for 14 out of 17 FGR fetuses below the 3rd percentile. Six fetuses referred to MRI as AGA were found to be < 3rd percentile.

CONCLUSIONS

The proposed method for automatic MRI-based EFW demonstrated high performance and sensitivity to identify FGR fetuses.

CLINICAL RELEVANCE STATEMENT

Results from this study support the use of the automatic fetal weight estimation method based on MRI for the assessment of fetal development and to detect fetuses at risk for growth restriction.

KEY POINTS

• An AI-based segmentation method with a quality assessment module for fetal weight estimation based on MRI was developed, achieving high repeatability, reproducibility, and accuracy. • An MRI-based fetal weight growth chart constructed from a large cohort of normal and appropriate gestational-age fetuses is proposed. • The method showed a high sensitivity for the diagnosis of small fetuses suspected of growth restriction.

Automatic Segmentation of the Fetus in 3D Magnetic Resonance Images Using Deep Learning: Accurate and Fast Fetal Volume Quantification for Clinical Use

Magnetic resonance imaging (MRI) provides images for estimating fetal volume and weight, but manual delineations are time consuming. The aims were to (1) validate an algorithm to automatically quantify fetal volume by MRI; (2) compare fetal weight by Hadlock's formulas to that of MRI; and (3) quantify fetal blood flow and index flow to fetal weight by MRI. Forty-two fetuses at 36 (29-39) weeks gestation underwent MRI. A neural network was trained to segment the fetus, with 20 datasets for training and validation, and 22 for testing. Hadlock's formulas 1-4 with biometric parameters from MRI were compared with weight by MRI. Blood flow was measured using phase-contrast MRI and indexed to fetal weight. Bland-Altman analysis assessed the agreement between automatic and manual fetal segmentation and the agreement between Hadlock's formulas and fetal segmentation for fetal weight. Bias and 95% limits of agreement were for automatic versus manual measurements 4.5 ± 351 ml (0.01% ± 11%), and for Hadlock 1-4 vs MRI 108 ± 435 g (3% ± 14%), 211 ± 468 g (7% ± 15%), 106 ± 425 g (4% ± 14%), and 179 ± 472 g (6% ± 15%), respectively. Umbilical venous flow was 406 (range 151-650) ml/min (indexed 162 (range 52-220) ml/min/kg), and descending aortic flow was 763 (range 481-1160) ml/min (indexed 276 (range 189-386) ml/min/kg). The automatic method showed good agreement with manual measurements and saves considerable analysis time. Hadlock 1-4 generally agree with MRI. This study also illustrates the confounding effects of fetal weight on absolute blood flow, and emphasizes the benefit of indexed measurements for physiological assessment.

Semiautomatic Assessment of Fetal Fractional Limb Volume for Weight Prediction in Clinical Praxis: How Does It Perform in Routine Use?

OBJECTIVES

Semiautomatic fractional limb volume (FLV) models have recently produced promising results for fetal birth weight (BW) estimation. We tested those models in a more unselected population hypothesizing that the FLV models would improve accuracy and precision of fetal BW estimation compared to the Hadlock model.

METHODS

We compared the performance of different BW prediction models: Hadlock (biparietal diameter [BPD], abdominal circumference (AC), femur diaphysis length) and modified Lee thigh volume (TVol) and arm volume (AVol) (BPD, AC, automated fractional TVol, and AVol). Accuracy (systematic errors, mean percent differences) and precision (random errors, ± 1 SD of percent differences) were calculated.

RESULTS

A total of 75 fetuses were included for final analysis. The Hadlock model showed the most consistent results with accurate BW estimation not significantly different from zero (-0.37 ± 8.53%). The modified fractional thigh and arm volume models were less accurate but trended toward more precise results (-2.63 ± 7.69% and -3.85 ± 7.47%, respectively). In addition, the modified TVol model showed the trend to predict more BWs within ±10% of the actual BW compared to the Hadlock model (81.3 versus 74.67%, ns).

CONCLUSIONS

Based on our results, fetal weight estimation using the modified semiautomatic FLV models generates less accurate results in third-trimester fetuses compared to the Hadlock model. Those models recently published might improve the results of BW prediction by showing a higher precision than conventional models, especially in small and large fetuses. Further studies are needed to investigate the clinical usefulness of the new models.

Impact of biometric measurement error on identification of small- and large-for-gestational-age fetuses

OBJECTIVES

First, to obtain measurement-error models for biometric measurements of fetal abdominal circumference (AC), head circumference (HC) and femur length (FL), and, second, to examine the impact of biometric measurement error on sonographic estimated fetal weight (EFW) and its effect on the prediction of small- (SGA) and large- (LGA) for-gestational-age fetuses with EFW < 10^th and > 90^th percentile, respectively.

METHODS

Measurement error standard deviations for fetal AC, HC and FL were obtained from a previous large study on fetal biometry utilizing a standardized measurement protocol and both qualitative and quantitative quality-control monitoring. Typical combinations of AC, HC and FL that gave EFW on the 10^th and 90^th percentiles were determined. A Monte-Carlo simulation study was carried out to examine the effect of measurement error on the classification of fetuses as having EFW above or below the 10^th and 90^th percentiles.

RESULTS

Errors were assumed to follow a Gaussian distribution with a mean of 0 mm and SDs, obtained from a previous well-conducted study, of 6.93 mm for AC, 5.15 mm for HC and 1.38 mm for FL. Assuming errors according to such distributions, when the 10^th and 90^th percentiles are used to screen for SGA and LGA fetuses, respectively, the detection rates would be 78.0% at false-positive rates of 4.7%. If the cut-offs were relaxed to the 30^th and 70^th percentiles, the detection rates would increase to 98.2%, but at false-positive rates of 24.2%. Assuming half of the spread in the error distribution, using the 10^th and 90^th percentiles to screen for SGA and LGA fetuses, respectively, the detection rates would be 86.6% at false-positive rates of 2.3%. If the cut-offs were relaxed to the 15^th and 85^th percentiles, respectively, the detection rates would increase to 97.0% and the false-positive rates would increase to 6.3%.

CONCLUSIONS

Measurement error in fetal biometry causes substantial error in EFW, resulting in misclassification of SGA and LGA fetuses. The extent to which improvement can be achieved through effective quality assurance remains to be seen but, as a first step, it is important for practitioners to understand how biometric measurement error impacts the prediction of SGA and LGA fetuses. © 2019 The Authors. Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of the International Society of Ultrasound in Obstetrics and Gynecology.

Ultrasonographic estimation of fetal weight: development of new model and assessment of performance of previous models

OBJECTIVES

To develop a new formula for ultrasonographic estimation of fetal weight and evaluate the accuracy of this and all previous formulae in the prediction of birth weight.

METHODS

The study population consisted of 5163 singleton pregnancies with fetal biometry at 22-43 weeks' gestation and live birth of a phenotypically normal neonate within 2 days of the ultrasound examination. Multivariable fractional polynomial analysis was used to determine the combination of variables that provided the best-fitting models for estimated fetal weight (EFW). A systematic review was also carried out of articles reporting formulae for EFW and comparing EFW to actual birth weight. The accuracy of each model for EFW was assessed by comparing mean percentage error, absolute mean error (AE), proportion of pregnancies with AE ≤ 10% and Euclidean distance.

RESULTS

The most accurate models, with the lowest Euclidean distance and highest proportion of AE ≤ 10%, were provided by the formulae incorporating ≥ 3 rather than < 3 biometrical measurements. The systematic review identified 45 studies describing a total of 70 models for EFW by various combinations of measurements of fetal head circumference (HC), biparietal diameter, femur length (FL) and abdominal circumference (AC). The most accurate model with the lowest Euclidean distance and highest proportion of AE ≤ 10% was provided by the formula of Hadlock et al., published in 1985, which incorporated measurements of HC, AC and FL; there was a highly significant linear association between EFW and birth weight (r = 0.959; P < 0.0001), and EFW was within 10% of birth weight in 80% of cases. The performance of the best model developed in this study, utilizing HC, AC and FL, was very similar to that of Hadlock et al. CONCLUSION: Despite many efforts to develop new models for EFW, the one reported in 1985 by Hadlock et al., from measurements of HC, AC and FL, provides the most accurate prediction of birth weight and can be used for assessment of all babies, including those suspected to be either small or large. Copyright © 2018 ISUOG. Published by John Wiley & Sons Ltd.

A comparison of ultrasound with magnetic resonance imaging in the assessment of fetal biometry and weight in the second trimester of pregnancy: An observer agreement and variability study

Objective

To compare the intra and interobserver variability of ultrasound and magnetic resonance imaging in the assessment of common fetal biometry and estimated fetal weight in the second trimester.

Methods

Retrospective measurements on preselected image planes were performed independently by two pairs of observers for contemporaneous ultrasound and magnetic resonance imaging studies of the same fetus. Four common fetal measurements (biparietal diameter, head circumference, abdominal circumference and femur length) and an estimated fetal weight were analysed for 44 'low risk' cases. Comparisons included, intra-class correlation coefficients, systematic error in the mean differences and the random error.

Results

The ultrasound inter- and intraobserver agreements for ultrasound were good, except intraobserver abdominal circumference (intra-class correlation coefficient = 0.880, poor), significant increases in error was seen with larger abdominal circumference sizes. Magnetic resonance imaging produced good/excellent intraobserver agreement with higher intra-class correlation coefficients than ultrasound. Good interobserver agreement was found for both modalities except for the biparietal diameter (magnetic resonance imaging intra-class correlation coefficient = 0.942, moderate). Systematic errors between modalities were seen for the biparietal diameter, femur length and estimated fetal weight (mean percentage error = +2.5%, -5.4% and -8.7%, respectively, p < 0.05). Random error was above 5% for ultrasound intraobserver abdominal circumference, femur length and estimated fetal weight and magnetic resonance imaging interobserver biparietal diameter, abdominal circumference, femur length and estimated fetal weight (magnetic resonance imaging estimated fetal weight error >10%).

Conclusion

Ultrasound remains the modality of choice when estimating fetal weight, however with increasing application of fetal magnetic resonance imaging a method of assessing fetal weight is desirable. Both methods are subject to random error and operator dependence. Assessment of calliper placement variations may be an objective method detecting larger than expected errors in fetal measurements.

Automated Fractional Limb Volume Measurements Improve the Precision of Birth Weight Predictions in Late Third-Trimester Fetuses

OBJECTIVES

Fetal soft tissue can be assessed by using fractional limb volume as a proxy for in utero nutritional status. We investigated automated fractional limb volume for rapid estimate fetal weight assessment.

METHODS

Pregnant women were prospectively scanned for 2- and 3-dimensional fetal biometric measurements within 4 days of delivery. Performance of birth weight prediction models was compared: (1) Hadlock (Am J Obstet Gynecol 1985; 151:333-337; biparietal diameter, abdominal circumference, and femur diaphysis length); and (2) Lee (Ultrasound Obstet Gynecol 2009; 34:556-565; biparietal diameter, abdominal circumference, and automated fractional limb volume). Percent differences were calculated: [(estimated birth weight - actual birth weight) ÷ (actual birth weight] × 100. Systematic errors (accuracy) were summarized as signed mean percent differences. Random errors (precision) were calculated as ± 1 SD of percent differences.

RESULTS

Fifty neonates were delivered at 39.4 weeks' gestation. The Hadlock model generated the most accurate birth weight (0.31%) with a mean random error of ±7.9%. Despite systematic underestimations, the most precise results occurred with fractional arm volume (-9.1% ± 5.1%) and fractional thigh (-5.2% ± 5.2%) models. The size and distribution of these prediction errors were improved after correction for systematic errors.

CONCLUSIONS

Automated fractional limb volume measurements can improve the precision of weight predictions in third-trimester fetuses. Correction factors may be necessary to adjust underestimated systematic errors when using automated fractional limb volume with prediction models that are based on manual tracing of fetal limb soft tissue borders.

Fetal weight estimation in gestational diabetic pregnancies: comparison between conventional and three-dimensional fractional thigh volume methods using gestation-adjusted projection

OBJECTIVES

To evaluate the accuracy of gestation-adjusted birth-weight estimation using a three-dimensional (3D) fractional thigh volume (TVol) method in pregnant women with gestational diabetes mellitus (GDM), and to compare it with the conventional two-dimensional method of Hadlock et al.

METHODS

Pregnant women with GDM were referred at 34 to 36 + 6 weeks' gestation for ultrasound examination. Estimated fetal weight (EFW) was obtained using both the Hadlock and the TVol methods. Using a gestation-adjusted projection method, predicted birth weight was compared to actual birth weight at delivery.

RESULTS

Based on 125 pregnancies, the TVol method with gestation-adjusted projection had a mean (± SD) percentage error in estimating birth weight of -0.01 ± 5.0 (95% CI, -0.96 to 0.98)% while the method of Hadlock with gestation-adjusted projection had an error of 1.28 ± 9.1 (95% CI, -0.33 to 2.87)%. The mean percentage error of the two methods was significantly different (P = 0.039), while the random error was not (P = 1.0). For the prediction of macrosomia (birth weight ≥ 4000 g, n = 19), sensitivity was 84 and 63% for the TVol and Hadlock methods, respectively (95% CI for difference -2 to 44%, P = 0.22) and specificity was 96 and 89% for the TVol and Hadlock methods, respectively (95% CI for difference 5-9%, P = 0.01).

CONCLUSIONS

In women with GDM, a new method of estimating birth weight based on 3D-TVol measurements performed at 34 + 0 to 36 + 6 weeks' gestation and gestation-adjusted projection of estimated fetal weight, is more accurate than the standard method based on Hadlock's formula in predicting birth weight. The TVol method has comparable sensitivity but higher specificity than the Hadlock method in predicting neonatal macrosomia.

What are the limits of accuracy in fetal weight estimation with conventional biometry in two-dimensional ultrasound? A novel postpartum study

OBJECTIVE

Commonly used formulae for fetal weight estimation, including combinations of several biometric parameters, lack accuracy despite efforts to improve them. This study aimed to investigate the limits of fetal weight estimation based on conventional biometric parameters on two-dimensional (2D) ultrasound by developing and evaluating new weight equations using postpartum biometric parameters.

METHODS

This was a prospective multicenter study including 628 singleton pregnancies at term. Inclusion criteria were healthy newborns with no physical or chromosomal malformations. Postpartum measurement of head circumference, abdominal circumference and thigh length was performed. Six 'best-fit' formulae were derived using forward regression analysis in a formula-finding group (n = 419), and their accuracy was compared with birth weight in an evaluation group (n = 209) using percentage error, absolute percentage error, limits of agreement and the proportion of weight estimations falling within a discrepancy level of ± 10%.

RESULTS

The new formulae showed no systematic error, with SD for the percentage error between 7.42 and 8.77 and no significant differences between median absolute percentage errors (4.84-5.71). They included 74.6-81.3% of neonates within a discrepancy level of 10%. With regard to the 95% limits of agreement, weight estimates were within a range of about ± 500 g.

CONCLUSION

These results show that a good sonographic weight formula has the following features: no systematic error, an SD of about 7% and inclusion of 80% of cases within a discrepancy level of 10%. The study indicates that the current accuracy of fetal weight estimation with conventional biometric parameters by 2D ultrasound has reached its limits. Further improvement will probably only be achieved through new approaches in ultrasonography.

Is sonographic assessment of fetal weight influenced by formula selection?

OBJECTIVE

Several published formulas exist for the determination of estimated fetal weight (EFW), with limited data on their comparative accuracies. The aims of our study were to assess and compare the performance of different EFW formulas in predicting actual birth weight (BW) in an urban population.

METHODS

Patients with an EFW determined within 7 days of delivery were considered eligible for the study. Fourteen published formulas, derived from populations comparable to ours, were used to recalculate EFWs from the same initial measurements. The accuracy of the EFWs obtained from the different formulas were compared by percentage error methods using bias and precision and Bland-Altman limits of agreement methods. Sensitivity and specificity for prediction of being small for gestational age (SGA) were calculated.

RESULTS

Eighty-one fetuses were included in the study. Formula C of Hadlock et al [Hadlock C; log(10) BW = 1.335 - 0.0034(abdominal circumference [AC])(femur length [FL]) + 0.0316(biparietal diameter) + 0.0457(AC) + 0.1623(FL); Am J Obstet Gynecol 1985; 151:333-337] had the best performance according to the bias and precision method. Bland-Altman limits of agreement confirmed these results. Among the formulas, the sensitivity for detection of SGA ranged from 72% to 100%, and specificity was 41% to 88%. Hadlock C had the optimal sensitivity/specificity trade-off for detection of SGA.

CONCLUSIONS

Fourteen formulas showed considerable variation of bias and precision in our population as well as a wide range of sensitivities and specificities for SGA. The choice of the appropriate formula for EFW in a given population should be based on objective and explicit criteria. Consideration of bias and precision for the formula in the population being assessed is critical and may affect clinical care.