Three-dimensional (3D) ultrasound is useful in the prenatal evaluation of fetal craniofacial structures, particularly as it provides a multiplanar view. However, an expert must designate the area of interest and the appropriate view, making measurement of fetal structures using 3D ultrasound both time-consuming and subjective. In this study we propose an image analysis system that measures automatically and precisely the fetal craniofacial structures and evaluate its performance in the second trimester of pregnancy using a new 3D volume analysis algorithm.
A universal facial surface template model containing the geometric shape information of a fetal craniofacial structure was constructed from a fetal phantom. Using the proposed image analysis system we fitted this stored template model using a model deformation approach to individual fetal 3D facial volumes from 11 mid-trimester fetuses, and extracted automatically the following standard measurements: biparietal diameter (BPD), occipitofrontal diameter (OFD), interorbital diameter (IOD), bilateral orbital diameter (BOD) and distance between vertex and nasion (VN). The same five parameters were measured manually by an expert and the results compared.
Comparison of the algorithm-based automatic measurements with manual measurements made by an expert gave correlation coefficients of 0.99 for BPD, 0.98 for OFD, 0.80 for BOD, 0.83 for IOD and 0.99 for VN. There were no significant differences between automatic and manual measurements.
Craniofacial dysmorphology with clinical implications has been described in over 150 syndromes, making it the keystone of syndrome delineation1. Minor dysmorphism can be difficult to assess and often depends on the clinical impression of the physician2. Ultrasound is the most convenient and powerful tool for evaluating fetal morphology in utero. Although several studies have attempted to improve the prenatal diagnosis of craniofacial malformations by describing normal craniofacial anatomy, minor abnormalities are difficult to determine using conventional two-dimensional (2D) ultrasound. Three-dimensional (3D) ultrasound has potential advantages, especially in the evaluation of complex anatomy, such as the fetal head and face3, 4. However, although 3D ultrasound provides a multiplanar view, an expert must designate manually an area of interest and an appropriate viewing plane in order to obtain precise measurement of fetal morphology. The decision regarding plane selection, which directly influences measurement accuracy, is usually time-consuming and subjective5.
In recent years, research investigating image analysis and pattern recognition algorithms for the purpose of image-guided diagnosis and surgery has indicated that the use of automation in medicine and its application to ultrasound is logical and feasible6. Kaus et al.7 proposed an automatic method for the construction of a 3D point distribution model from segmented images: given a set of segmented objects in a volumetric image, they selected an arbitrary template and triangulated the object surface as a template mesh. For each of the objects, the template mesh was adapted locally to the object shape using a deformable model-based approach. Ecabert et al.8 built a statistical model of the human heart using a similar approach. Furthermore, Yu et al.9 proposed a four-step image segmentation technique for fetal abdominal contour extraction and measurement from 2D ultrasound images.
In this study, employing image analysis techniques10–14, we propose and implement an automatic system for measuring fetal craniofacial structures from 3D ultrasound volumes, validating its performance using 11 3D fetal ultrasound images.
Subjects and Methods
This study was approved by National Cheng Kung University Hospital Institutional Review Board and participants gave informed consent. Between January 2010 and December 2010, we recruited from the antenatal outpatient department of National Cheng Kung University Hospital 11 women without disease known to affect fetal growth, i.e. pre-existing hypertension or diabetes mellitus, and with normal singleton pregnancies not at risk for craniofacial abnormality. Pregnancy duration was determined from the last reliable menstrual period or, in case of uncertainty, it was adjusted by ultrasound in the early first trimester.
During the routine anomaly scan, a single operator (Y.-C.C.) acquired 11 static volumes from the fetuses, at confirmed gestational ages of 20–24 weeks, using a transabdominal 3D transducer with a frequency range of 4–8 MHz (Voluson 530 D or Voluson 730 Expert, GE Medical Systems, Zipf, Austria). The acquisition angle, which was 85° for the majority of volume measurements, was set to ensure inclusion of the fetal head vertex superiorly and the fetal upper neck region inferiorly. The borders of the acquisition box were placed outside the fetal head. Volumes were stored for later offline analysis.
Automatic measurement of fetal craniofacial structure
Five craniofacial measurements (biparietal diameter (BPD), occipitofrontal diameter (OFD), interorbital diameter (IOD), bilateral orbital diameter (BOD) and distance from vertex to nasion (VN)), which have been proved useful in the assessment of prenatal15 and postnatal16–20 normal craniofacial development, were extracted under our proposed system using anthropometric (soft-tissue landmarks) and cephalometric (radiographically derived or bony landmarks) landmarks as reported in previous studies15–20. This was achieved using a model-based segmentation method to delineate automatically the fetal craniofacial structure, as described below.
The model used in the algorithm was constructed from a fetal ultrasound training phantom (CIRS model 065-20, Computerized Imaging Reference Systems, Norfolk, VA, USA), with a gestational age of 20 weeks, suspended in an anechoic and amniotic fluid-like environment. A volumetric ultrasound image of the phantom was segmented using fundamental image processing techniques, including smoothing filter, thresholding, and region growing21, and then triangulated with the marching cube algorithm10 to obtain a reference model (Figure 1). The soft tissue and bony landmarks for determining the craniofacial measurements were then identified manually by an expert obstetrician (P.-Y. T.) on the model surface.
To obtain measurements from the 11 fetal 3D ultrasound images, any differences in pose (i.e. position, orientation and size) (Figure 2a) between the model and each fetus had to be eliminated via feature-based registration. The basic concept of registration was to find a geometric transformation which could adjust the pose of the model to accommodate the ultrasound volume. The feature points used in this registration process were located in the central region of the head and in the eyes. The least squares ellipse fitting algorithm11 was used to outline the head region and the Gabor texture12 was utilized to detect the eyes. Corresponding feature points between the model and the fetal image were aligned13, so that the pose of the registered reference model matched that of the ultrasound image (Figure 2b), on visual comparison with before registration (Figure 2a). A 3D snake-based algorithm was then used to refine the reference model's surface14; each vertex of the model was moved iteratively towards the boundary of the fetal head, which was defined in the image by the transition from high to low intensity (Figure 2c). Since this deformation process did not change the global topology of the model's surface, the landmark locations in the 3D ultrasound image could be identified readily, based on the expert-defined landmarks on the reference model. The craniofacial measurement results were obtained automatically as distances between these landmark locations. On every 3D volume, the five measurements were each performed five times automatically by the algorithm.
In order to validate the performance of the proposed system, an expert obstetrician (P.-Y.T.), who was blinded to the automatic measurement results, measured manually five times each of the same five craniofacial parameters from all 11 fetal ultrasound volumes, for comparison with the results of the automatic segmentation method.
We estimated the variance between automatic and manual measurements. Five replicates for each measurement were used to calculate the mean and variance. The results are presented as mean, SD and 95% CIs of the difference between automatic and manual measurements. The relationship between automatic and manual measurements was assessed by a paired sample t-test, and P < 0.05 was considered statistically significant. Statistical analysis was carried out using the Statistical Package for the Social Sciences (SPSS 17.0 for Windows, SPSS Inc., Chicago, IL, USA). Bland–Altman plots were used to assess bias. The contours obtained as part of the automatic measurement process were also assessed by comparison with contours obtained manually by outlining the 2D ultrasound images (P.-Y.T.) from the volume data. As it would be impractical to manually trace all contours in a 3D volume (there being around 100 images per volume), we used an image close to the midsa gittal plane of the fetal head, measuring the distance from each pixel in the automatically obtained contour to the nearest pixel in the manually obtained contour and calculating their mean.
Comparison of automatic measurements from the 11 static 3D volumes with the dimensions obtained manually by an expert showed no significant differences and obtained high correlation coefficients for BPD, OFD, IOD, BOD and VN (Table 1), confirming that the automatic method achieved measurement results consistent with those obtained manually by an expert. Differences between the automatic and manual measurement results were examined using Bland–Altman plots (Figure 3); results obtained using the proposed automatic method showed good agreement with those obtained using the manual method. The difference in contours between the two methods was less than 0.6 mm in mean distance (Figure 4).
Table 1. Comparison of automatic and manual methods in fetal craniofacial indices from 11 fetal ultrasound volumes
95% CI of difference
Voxel, volumetric pixel or volumetric picture element; each voxel represents 0.58 mm in physical space.
Bilateral orbital diameter
Due to its relatively low cost, real-time results and non-invasive nature, ultrasound is the most convenient and powerful tool for the evaluation of fetal growth in utero. Developing a reliable measurement system is desirable for facilitating diagnosis and treatment. Automatic assisted systems generally increase efficiency, reliability and accuracy and/or minimize cost22.
In traditional 2D ultrasound, finding the optimal plane is a major requirement for obtaining valid, precise and reproducible measurements and achieving good quality of examination23, 24. However, this skill may require a lot of expert training. The proposed 3D ultrasound image analysis technique allows simultaneous display of three perpendicular planes in the multiplanar view. The biometry planes and craniofacial landmarks which are critical to traditional 2D measurement are provided automatically by the proposed system, through comparison to the reference model and its predefined, known landmarks.
Craniofacial structural measurement has been proved useful in the assessment of normal craniofacial development and of congenital fetal anomalies. In this study, we used five measurements (BPD, OFD, IOD, BOD, VN) shown to be clinically useful in a previous study19. These measurements describe the facial width, depth and height from anthropometric and cephalometric measurements15–20 and were the initial measurements selected for use with our newly created automatic system because they are easy to identify and show clear definition on 3D ultrasound. In future, we will attempt to evaluate more fetal facial parameters, such as the maxilla–nasion–mandible angle25.
Increasing numbers of semi-automatic and automatic systems have been developed for sonographic fetal evaluation. Moratalla et al.26 and Abele et al.27 reported that semi-automated measurement of fetal nuchal translucency thickness may be a useful alternative to manual measurement. 3D ultrasound software has been applied to fetal hearts to detect the planes and volumes of interest28, 29. However, there is a lack of computerized image analysis systems for automatic evaluation of the fetal craniofacial structure in the second trimester of pregnancy. In the present study, therefore, we developed an automatic digital system to help resolve the problems encountered in using conventional 2D manual methods5.
Our proposed image processing system measures the craniofacial parameters directly from the 3D image data, and thus can obtain precise measurements of fetal cranial structure without the influence of having to select the best 2D view. The results presented here validate the automated measurement of the fetal craniofacial structure using 3D ultrasound and provide evidence for its potential clinical applicability. Our proposed system could obtain automatically, without user intervention, all five measurements in around 30 s; moreover, the resulting measurements were very close to corresponding manual measurements. We believe that the overall trade-off between time and accuracy is acceptable.
A limitation of the system is that when the fetus moves or the fetal head has other soft tissues adhering to it during 3D ultrasound acquisition, the image analysis becomes complicated, making it difficult to retrieve a complete set of measurements. Another is the difficulty of recording the entire fetal head at advanced gestational age due to the limited 3D transducer sector size; we therefore used gestational ages of 20–24 weeks in this study.
In the future, we plan to assess reproducibility and obtain normal growth data relative to gestational age for these measurements. After establishing the normal fetal craniofacial measurements using our proposed system, it will be easier to detect craniofacial dysmorphology. Moreover, we will conduct further analysis of other craniofacial patterns or profiles to assist the evaluation of fetal cranial anatomy. When implementing this analysis algorithm in a commercial system, we also plan to optimize it for the system program/code in order to maximize the system's performance. This system could be a valuable tool for delineating and distinguishing syndromes affecting craniofacial development. We believe that automatic detection and measurement of the fetal craniofacial structure is clinically useful, and that our proposed system may be usefully applied to other clinical fields in the future.
This study was supported by a grant from the National Science Council of the Republic of China (NSC 100-2314-B-006-013-MY3) and intramural grants from National Cheng-Kung University Hospital, Tainan, Taiwan. We are grateful to Ms Yu-Ting Wu for her assistance.