In literature there is little research available on automatic analysis of first trimester ultrasound in human pregnancies. Two studies focus on performing segmentation of the embryo in first trimester 3D ultrasound (Yang et al., 2018; Looney et al., 2021). Both approaches use a fully convolutional neural network, combined with transfer learning and either label refinement using a recurrent neural network (Yang et al., 2018) or a two-pathway architecture (Looney et al., 2021). They achieve mean Dice score of 0.876 (104 3D volumes, 10-14 weeks GA) and 0.880 (2393 volumes, 11-14 weeks GA) respectively. Carneiro et al. (2008) studied automatic extraction of the biparietal diameter, head circumference, abdominal circumference, femur length, humerus length, and CRL in first trimester ultrasound using a constrained probabilistic boosting tree classifier. Finally, three studies focused on obtaining the head circumference using publicly available data, which consists of 2D ultrasound scans of the trans-thalamic plane (van den Heuvel et al., 2018; Al-Bander et al., 2019; Li et al., 2020).

Limitation of the aforementioned studies is that they focus on extracting individual measurements or segmentations, whereas we propose to align the embryo to a predefined standard space. This is beneficial since alignment to a standard space allows for straightforward extraction of various biometric and volumetric measurements in a unified framework.

Most published studies focus on performing spatial alignment and segmentation separately; for a comprehensive overview see Torrents-Barrena et al. (2019). When both spatial alignment and segmentation are of interest performing them sequentially raises the question of the optimal order of operations. Especially since some studies focusing on performing segmentation require prior knowledge about spatial orientation of the embryo (Gutiérrez-Becker et al., 2013; Yaqub et al., 2013) and some studies focusing on standard plane detection require the segmentation or detection of structures within the embryo (Ryou et al., 2016; Yaqub et al., 2015). To circumvent this, a few studies perform both tasks at once (Chen et al., 2012; Kuklisova-Murgasova et al., 2013; Namburete et al., 2018).

Chen et al. (2012) achieved both tasks for the fetal head using spatial features from eye detection and image registration techniques, for ultrasound images acquired at 19-22 weeks GA. However, the eye is not always clearly detectable in 3D first trimester ultrasound, making this method not applicable in our case.

Kuklisova-Murgasova et al. (2013) achieved both tasks for the fetal brain, using a magnetic resonance imaging (MRI) atlas and block matching using image registration techniques, for ultrasound images acquired at 23-28 weeks GA. However, such an atlas or reference MRI image is not available for the first trimester (Oishi et al., 2019).

Namburete et al. (2018) achieved both alignment and segmentation of the fetal head with a supervised multi-task deep learning approach, using slice-wise annotations containing a priori knowledge of the orientation of the head, and manual segmentations, for ultrasound images acquired at 22-30 weeks GA. However, this method assumes that during acquisition the ultrasound probe is positioned such that the head is imaged axially, which is not always the case for 3D first trimester ultrasound.

Although none of the studies are applicable to first trimester ultrasound, they all employ image registration techniques to perform segmentation and spatial alignment simultaneously. In line with this, we take an image registration approach and tailor our method to be applicable to first trimester ultrasound.

Recently, deep learning methods for image registration have been developed and show promising results in many areas; for an extensive overview see Boveiri et al. (2020). The common assumption for most of the deep learning approaches for image registration is that the data is already affinely registered. However, due to the rapid development of the embryo and the wide variation in position and spatial orientation the affine registration is challenging to obtain and therefore has to be incorporated in the framework.

There are a few studies focusing on achieving both affine and nonrigid registration. For example, de Vos et al. (2019), proposed a multi-stage framework, dedicating the first network to learning the affine transformation and the second network to learning the voxelwise displacement field. These networks are trained in stages. Zhengyang et al. (2019) proposed a similar approach where the main difference is that they used a diffeomorphic model for the nonrigid deformation. We also take a multi-stage approach with a dedicated network for the affine and nonrigid deformation. We based the design of both networks on Voxelmorph by Balakrishnan et al. (2019), which is developed for deep learning-based nonrigid image registration and is publicly available.

Using multiple atlases is a common and successful technique for biomedical image segmentation. Adding data of multiple atlases gives in general better and more robust results and is widely used in classical not learning-based methods (Iglesias and Sabuncu, 2015). However, learning-based methods can also benefit from using multiple atlases (Ding et al., 2019; Fang et al., 2019; Lee et al., 2019).

Fang et al. (2019) omits the registration step by directly guiding the training of a segmentation network using multiple atlases as ground truth. Ding et al. (2019) and Lee et al. (2019) both employ deep learning for the image registration step, but they use different strategies for training. Ding et al. (2019) trained a network that warps all available atlas to the target image at once. On the other hand, Lee et al. (2019) trains the network by giving at every epoch a random atlas to the network to register to the target image.

Both approaches train the network to be able to register all available atlases to all target images. In our case, the set of atlases consists of ultrasound images from different subjects acquired in every week GA. Due to the rapid development of the embryo, only atlases acquired at approximately the same GA as the image are useful for registration. Hence, we compared different fusion strategies that take this into account.

Preliminary results of this framework were published in Bastiaansen et al. (2020a) and Bastiaansen et al. (2020b). In Bastiaansen et al. (2020a) we proposed a fully unsupervised atlas-based registration framework using deep learning for the alignment of the embryonic brain acquired at 9 weeks GA. We showed that having a separate network for the affine and nonrigid deformations improved the results. Subsequently, in Bastiaansen et al. (2020b) our previous work was extended by adding minimal supervision using the crown and rump landmarks to align and segment the embryo.

Here, we substantially extend our previous work by making it applicable to data acquired between the 8th and 12th week of pregnancy. We achieve this by taking a multi-atlas approach with atlases consisting of longitudinally acquired ultrasound images from multiple pregnancies. We compare different fusion strategies to combine data from multiple subjects and address the influence of differences in image quality of the atlas.

Compared to our previous work, we performed a more in-depth hyperparameter search for our deep learning framework and improved the division of the data in training, validation and test set taking into account the distribution of GA and available information used for evaluation, such as embryonic volume. Furthermore, we compared our best model to the 3D nn-UNet (Isensee et al., 2021), which is a current state-of-the-art fully supervised segmentation method. Finally, to obtain the segmentation of the embryo, in the proposed model the inverse of the nonrigid deformation is needed. To make our deformations invertible, we adopt a diffeomorphic deformation model as proposed by Dalca et al. (2019), using the stationary velocity field.

We present a deep learning framework to simultaneously segment and spatially align the embryo in image $I$. We propose to achieve this by registering every image $I$ to an atlas via deformation $\phi$:

with $\Omega$ the 3D domain of $A,I$. We assume that the atlas $A$ is in a predefined standard orientation and the segmentation $S_{A}$ is available. Now, $I\circ\phi$ is in standard orientation and the segmentation of image $I$ is given by:

The deep learning framework learns to estimate deformation $\phi$ given atlas $A$ and the image $I$. We propose to use a multi-atlas approach by registering every image to the $M$ atlases that are closest in GA. In Fig. ​1 a detailed overview of our framework can be found; all components are explained in the remainder of this section.

The input of the affine registration network is the image $I$. The affine registration network gives as output the estimated affine transformation $\phi_{a}$ and the corresponding affinely registered image $I\circ\phi_{a}$. Recall that we affinely registered every atlas to the same standard orientation. Hence, the affine transformation $\phi_{a}$ registers the image $I$ to all atlases in $\mathcal{A}$. Furthermore, this allowed us to directly compare the image similarity of $I\circ\phi_{a}$ and all selected atlas images in the loss function. Hence, as shown by the arrows in Fig. ​1, in contrast to the nonrigid network, the atlas image $A_{i,j}$ is not given as input to the network.

The input of the nonrigid registration network is an affinely registered image $I_{\phi_{a}}:=I\circ\phi_{a}$ together with a selected atlas $A_{i,j}$. During training we provide as input to the network the image $I$ and one randomly chosen selected atlas $A_{i,j}$. During inference we give as input to the network all the possible pairs of image and selected atlas. The output of the network consists of $\phi_{d_{i,j}}$, along with the registered images $I_{\phi_{a}}\circ\phi_{d_{i,j}}$.

To obtain the segmentation in Eq. ​6, the deformation $\phi_{d_{i,j}}$ must be inverted. To ensure invertibility, we adopt diffeomorphic deformation for $\phi_{d_{i,j}}$. Following Dalca et al. (2019), we use a stationary velocity field (SVF) representation, meaning that $\phi_{d_{i,j}}$ is obtained by integrating the velocity field $\nu$ (Ashburner, 2007). Now, $\phi_{d_{i}}^{-1}$ is obtained by integrating the velocity field using $-\nu$.

For training the ADAM optimizer was used with a learning rate of $10^{-4}$, as in Balakrishnan et al. (2019). For the second training stage of the affine registration network, we lower the learning rate to $10^{-5}$, since the network has already largely converged in the first stage. We trained each part of the network for 300 epochs with a batch size of one. In the first training stage of the affine network data augmentation was applied. Each scan in the training set was either randomly flipped along one of the axes or rotated 90, 180 or 270 degrees. For the second stage and nonrigid network preliminary experiments showed that data augmentation did not improve the results. The framework is implemented using Keras (Chollet and et al., 2015) with Tensorflow as backend (Abadi et al., 2016). The code is publicly available at:
https://gitlab.com/radiology/prenatal-image-analysis/multi-atlas-seg-reg

The Rotterdam Periconceptional Cohort (Predict study) is a large hospital-based cohort study conducted at the Erasmus MC, University Medical Center Rotterdam, the Netherlands. This prospective cohort focuses on the relationships between periconceptional maternal and paternal health and embryonic and fetal growth and development (Rousian et al., 2021; Steegers-Theunissen et al., 2016). 3D ultrasound scans are acquired at multiple points in time during the first trimester. Ultrasound examinations were performed on a Voluson E8 or E10 (GE Healthcare, Austria) ultrasound machine. Furthermore, maternal and paternal food questionnaires are taken and other relevant data such as weight, height and bloodsamples are collected.

The following four metrics were used to evaluate the results. Firstly, we report the relative error between the volume calculated from estimated segmentation $S_{I}$, referred to as $EV(I)$, and the ground truth embryonic volume $EV_{\text{gt}}(I)$:

Secondly, we calculate the target registration error (TRE) in millimeters (mm) for the crown and rump landmarks as defined in the loss function in Eq. ​12, to evaluate the accuracy of the affine alignment. The TRE is not calculated to evaluate the accuracy of the non-rigid alignment, since the TRE is only calculated in two landmarks, which is too limited.

Thirdly, we report the Dice similarity coefficient and the mean surface distance (MSD) in mm between the available manual segmentations $S_{I}^{\text{gt}}$ and the estimated segmentation $S_{I}$ of the image $I$ (Dice and Lee, 1945).

To compare the results of different experiments, we performed the two-sided Wilcoxon signed-rank test with a significance level of 0.05, we report the p-values for the following three metrics as $p_{EV}$ and $p_{Dice}$ and $p_{MSD}$. We excluded the TRE from the statistical analysis, since the TRE only measures the alignment in two points, and therefore a lower TRE does not by definition means better alignment.

We performed the following seven experiments.

Experiment 1: sensitivity metrics
In literature there are no comparable studies available, to get insight in the performance of the presented framework we addressed the sensitivity to perturbations of the used evaluation metrics. We performed 4 operations on the 184 available ground truth segmentations: 1) dilation using a ball with radius of one voxel: $S_{gt}\oplus B_{1}$; 2) erosion using a ball with a radius of one voxel: $S_{gt}\ominus B_{1}$; 3) dilation using a ball with a radius of two voxels: $S_{gt}\oplus B_{2}$; 4) erosion using a ball with a radius of two voxels: $S_{gt}\ominus B_{2}$. We calculated for the four obtained segmentations the $\text{EV}_{\text{error}}$ and Dice score with respect to $S_{gt}$ and report the mean and standard deviation per operation.

Experiment 2: influence of hyperparameters and second affine training stage
Using the single-subject strategy $\mathcal{A}_{1}^{s}$ we performed a hyperparameter search on the validation set for $\lambda_{l}\in\{0.01,1,100\}$ in Eq. ​9, $\lambda_{s}\in\{0,0.005,0.05,0.5,5\}$ in Eq. ​13 and $\lambda_{d}\in\{1,10,100\}$ in Eq. ​15 and the number of filters $(f_{1},f_{2})$ with $f_{2}=2f_{1}$ and $f_{1}\in\{16,32,64\}$ in the encoder of the affine network and $(f_{3},f_{4})$ with $f_{4}=2f_{3}$ and $f_{3}\in\{1,2,4,8,16\}$ in the encoder and decoder of the nonrigid network. Furthermore, we took out training stage 2 of the affine network to evaluate its necessity. We call this strategy $\mathcal{A}_{1}^{s}$[no stage 2].

Experiment 3: single-subject strategy
We used the best hyperparameters from experiment 2 to train $\mathcal{A}_{k}^{s}$ for $k=2,3,...,8$ and compare the results on the validation set to the results for $\mathcal{A}^{s}_{1}$, to evaluate whether the choice of subject $k$ influences the results. Finally, we applied $\mathcal{A}^{s}_{k}$ for best performing subject $k$ to the test set.

Experiment 4: comparison of the multi-subject and ensemble strategy
To evaluate what is the best approach to combine data of multiple subjects we compared the results on the validation set of the multi-subject strategy $\mathcal{A}^{m}$ for $M=1$, $M=2$, $M=4$ and $M=8$ to the ensemble strategy $\mathcal{A}^{e}$.

Experiment 5: influence of atlas quality
We investigated the influence of the quality of the ultrasound images on the results by excluding subject $i=8$, since this subject had the lowest image quality score (see Fig. ​3). We repeated experiment 4 without subject $i=8$ on the validation set and compared the results to the original results in experiment 4.

Experiment 6: comparison of the best models
We compared the best strategy using data of multiple subjects to the best single-subject strategy on the test set. This was done to investigate whether utilizing data of multiple subjects improves the results. In addition, for the best performing model, we compared the results to the 3D nn-UNet (Isensee et al., 2021). The nn-UNet was trained using: 1) the available ground truth segmentations of all atlases, and 2) additionally using the ground truth segmentations available from the validation set. This resulted in a training set of 40 and respectively 130 images. The nn-UNet was trained using the default parameters.

Experiment 7: analysis of best model
In the last experiment we evaluated the best performing strategy for the proposed method in more detail. Recall from Tab. ​1 that in our dataset, for every subject at one or more weeks GA, one or more ultrasound images were acquired. These images vary in quality, and in orientation and position of the embryo. In clinical practice, only one of the acquired images is used for the measurement, hence the algorithm has to work for at least one of the available ultrasound images. Firstly, we visually scored the affine alignment for every ultrasound image as follows: 0: poor, 1: landmarks aligned, 2: acceptable, 3: excellent. For the further analysis we took per subject and week GA the best score. Secondly, we plotted for cases with scores 2 and 3 the obtained embryonic volume against the ground truth embryonic volume, to compare their distribution and calculate the correlation between them. Thirdly, we show a visualization of the resulting mid-sagittal after alignment and the obtained segmentation. Fourthly, to gain more insight into reasons why our framework might fail, we visually inspected all images in the test set with a score of 0.

We performed four operations ($S_{gt}\oplus B_{1}$, $S_{gt}\ominus B_{1}$, $S_{gt}\oplus B_{2}$, $S_{gt}\ominus B_{2}$) on the 184 available ground truth segmentations to investigate the sensitivity to perturbations. From Tab. ​2(b) we conclude that an under- of overestimation of one voxel around the edge results in a Dice score between 0.81 and 0.84 on average and a $\text{EV}_{\text{error}}$ between 0.31 and 0.37 on average. An over- or underestimation of two voxels around the edge results in a Dice score between 0.61 and 0.71 and an $\text{EV}_{\text{error}}$ between 0.55 and 0.83. In the remaining experiments, these values serve as a reference for the residual errors.

In Fig. ​4 the mean $\text{EV}_{\text{error}}$, Dice score, TRE and MSD can be found for the validation set for different hyperparameters. For the first stage of the affine registration network, the best results were found for $\lambda_{l}=1$ and $(f_{1},f_{2})=(32,64)$. Setting $\lambda_{l}=0.01$ gave worse results, since the the supervision by landmarks was neglected. For $\lambda_{l}=100$ the results are worse, since in this case the image similarity term in Eq. ​(9) is neglected in the optimization.

Furthermore, Fig. ​4 shows that after applying the second training stage of the affine network in general the results improved. Only for $\lambda_{s}=5$ the results deteriorate, which is caused by penalizing scaling too much. The best results were found for $\lambda_{s}=0.05$. Note that the TRE increased a bit during the second training stage, which was to be expected, since the manually annotated landmarks have a limited precision and only measure alignment at two positions. Furthermore, we took $\lambda_{s}=0.05$ as optimal value, although for $\lambda_{s}=\{0,0.005\}$ the TRE is lower. This choice was made since the other metrics are optimal for $\lambda_{s}=0.05$, which indicates that the alignment in the two landmarks can be better with more scaling, but the overall quality deteriorates.

Regarding the results of training the nonrigid registration network, Fig. ​4 shows that setting $\lambda_{d}=100$ results in low flexibility of the deformation field, which leads to no improvements of the results after nonrigid registration. Setting $\lambda_{d}=1$ resulted in more irregular deformation fields, which led to less improvement of the metrics compared to $\lambda_{d}=10$. The best results were found for $\lambda_{d}=10$ and $(f_{3},f_{4})=(8,16)$.

Finally, we compared using the results of stage 2 ($\mathcal{A}_{1}^{s}$) or the results of stage 1
($\mathcal{A}_{1}^{s}$[no stage 2]) as input for the nonrigid registration network. Comparing the results for $\mathcal{A}_{1}^{s}$ and $\mathcal{A}^{s}_{1}$[no stage 2] in Fig. ​5 shows that the results are better for all metrics for $\mathcal{A}_{1}^{s}$. The statistical test confirmed these observations ($p_{EV}<0.001$, $p_{Dice}<0.001$, $p_{MSD}<0.001$).

Using the best hyperparameters determined in experiment 2, we trained the framework using the single-subject strategy for every subject $k=1,...,8$.Fig. ​5 shows for the validation set that for $\mathcal{A}_{k}^{s}$, for $k=1,..,8$ the results improved between the stages of the affine network and after nonrigid registration in terms of the $\text{EV}_{\text{error}}$, Dice score and MSD. For the TRE we observed again in general a slight increase after the second training stage. Furthermore, a wide spread is observed in the results, which comes from cases where the affine registration fails and this can not be compensated for by the subsequent nonrigid registration network.

In general our approach gives similar results independent of the choice of subject. However, there are differences in results between the different subjects, which are consistent with the variation in atlas image quality, since for example subject 8 had the least overall quality. The atlas image quality is given in Fig. ​3.

We applied the framework on the test set using the best hyperparameters found and the best single-subject strategy. Fig. ​5 shows that $\mathcal{A}_{4}^{s}$ performs the best in terms of Dice score and TRE and comparable in terms of $\text{EV}_{\text{error}}$ and MSD. When applying this model to the test set we found that the results in Fig. ​5 for $\mathcal{A}^{s}_{4}$[test] are comparable to those on the validation set ($\mathcal{A}^{s}_{4}$).

In experiment 4 we compared the multi-subject strategy $\mathcal{A}^{m}$, for different values of $M$, with the ensemble strategy $\mathcal{A}^{e}$. Fig. ​6a shows that for the $\text{EV}_{\text{error}}$ using the multi-subject strategy $\mathcal{A}^{m}$ gave for all tested values of $M$ better results than taking the ensemble strategy $\mathcal{A}^{e}$. This is confirmed by the statistical tests, with the p-values in Tab. ​3.

For the Dice score and MSD the ensemble strategy gave comparable results for the best multi-subject strategies with $M=\{4,8\}$.

For the $TRE$ we observe in Fig. ​6a that the results for the ensemble strategy $\mathcal{A}^{e}$ are better than for the multi-atlas strategy $\mathcal{A}^{m}$. This was to be expected since the mean of $n_{p}=8$ affine transformations is taken, which reduces the error in the alignment. However, despite the better affine alignment, the segmentation did not improve. Hence, we conclude that the multi-subject strategy performs better.

Regarding the value of $M$ for the multi-subject strategy, we found that the results for $M=\{4,8\}$ significantly outperformed the results for $M=\{1,2\}$ for at least one metric. When comparing $M=4$ to $M=8$, we found that $M=4$ had a lower median value for the $\text{EV}_{\text{error}}$ (0.31 versus 0.33) and MSD (1.61 versus 1.69), and higher median value for the Dice score (0.73 versus 0.70). Hence we conclude that the multi-subject approach with $M=4$ is the best strategy to include data from multiple subject.

After excluding subject $i=8$ with the lowest image quality, experiment 4 was repeated for the other $n_{p}=7$ subjects. Fig. ​6b shows that for multi-pregnancy strategy $\mathcal{A}^{m}$ with $M=1$ and the ensemble strategy $\mathcal{A}^{e}$ the results are significantly better when excluding the subject with lowest image quality ($\mathcal{A}^{m}$, $M=1$: $p_{EV}=0.001$, $p_{Dice}<0.001$, $p_{MSD}<0.001$, $\mathcal{A}^{e}$: $p_{EV}<0.001$, $p_{Dice}<0.001$, $p_{MSD}<0.001$). This can be explained by the fact that if fewer atlases are taken into account, atlases with worse quality will have more influence.

When comparing the multi-subject strategy $\mathcal{A}^{m}$ with $M=7$ against the multi-subject strategy $\mathcal{A}^{m}$ with $M=8$, the $\text{EV}_{\text{error}}$, and was significantly better for $M=7$ ($p_{EV}=0.011$, $p_{Dice}=0.434$, $P_{MSD}=0.510$). For the multi-subject strategy with $M=2$ and $M=4$ the results were not significantly different for any metric ($M=2$: $p_{EV}=0.229$, $p_{Dice}=0.121$, $p_{MSD}=0.291$, $M=4$: $p_{EV}=0.404$, $p_{Dice}=0.445$, $p_{MSD}=0.741$).

Therefor, since the multi-subject strategy was not harmed by the inclusion of the 8th atlas subject with the worst quality, we conclude that the multi-subject strategy $A^{m}$, with $M=2$ or $M=4$ is more robust to choice of atlas.

The results for the best single-subject strategy $\mathcal{A}^{s}_{4}$ and best multi-subject strategy $\mathcal{A}^{m}$ with $M=4$ were compared on the test set and no significant differences were found ($p_{EV}=0.75$, $p_{Dice}=0.71$, $p_{MSD}=0.86$). From the results we conclude that the multi-subject strategy with $M=4$ gives the best model, since this strategy omits choosing the right subjects as atlas. Furthermore, in experiment 5 we saw that the results are not affected by the presence of atlases with lower image quality. Finally, on the test set we find a median $\text{EV}_{\text{error}}$ of 0.29, Dice score of 0.72 and MSD of 1.58 mm.

Fig. ​7 shows the results of the comparison with the 3D nn-UNet. The nn-UNet is trained on 1) the atlas subjects, 2) the atlas subjects plus images with a ground truth from the validation set. For every week GA, the nn-UNet outperformed our approach for segmentation.

We analyzed the results using the multi-subject strategy $\mathcal{A}^{m}$ with $M=4$ in more detail for the test set. Measuring the embryonic volume semi-automatically in VR takes between 5 and 10 minutes (Rousian et al., 2013) per 3D image, the proposed framework takes on an Intel(R) Core(TM) i7-6850K CPU @ 3.60GHz 16 seconds to process one 3D image. Hence, we conclude that with appropriate hard- and software optimization our method can be used real-time in clinical practice.

In Tab. ​4 the results for the visual scoring are given. Overall, we found the following distribution: 0 (poor): 0.13 $\%$, 1 (landmarks aligned): 26 $\%$, 2 (acceptable): $39\%$, 3 (excellent): $21\%$. In Fig. ​9 examples for every score are given.

Next, we analyzed the metrics per score in Fig. ​8. For each metric we observed that in general a better visual score corresponds to a lower $\text{EV}_{\text{error}}$, TRE, MSD, and a higher Dice score. Furthermore, we observe that the wide spread in the results for all metrics is many caused by failed affine alignment (score 0 and 1), and that in cases with successful alignment (score 2 and 3) the spread for the TRE, MSD and Dice score is smaller. For the $\text{EV}_{\text{error}}$ for scores 2 and 3 there is a wide spread, however this spread is comparable to the values found in experiment 1 and for the nn-UNet in experiment 6.

Furthermore, we observed that all metrics are in the same range regardless of the week GA. Except for the $\text{EV}_{\text{error}}$ in week 8 and TRE in week 12. From the visual scoring in Tab. 4 we found that week 8 has relative the most poorly aligned (score 0) cases, and this was reflected in the $\text{EV}_{\text{error}}$. For week 12 the TRE became the largest in case of poor alignment, which can be explained by the the larger size of the embryo in week 12.