Two-dimensional ultrasound measurements vs. magnetic resonance imaging-derived ventricular volume of preterm infants with germinal matrix intraventricular haemorrhage

Very-low-birth-weight (<1,500 g) infants are at increased risk of developing a germinal matrix intraventricular haemorrhage, which occurs in 20–25% of cases [1, 2]. Approximately 35% of the infants who have a germinal matrix intraventricular haemorrhage develop post-haemorrhagic ventricular dilatation [3]. In severe cases of post-haemorrhagic ventricular dilatation, there is an increased risk of raised intracranial pressure and subsequent brain damage, and an intervention such as a lumbar puncture, the placement of a subcutaneous reservoir or a ventriculo-peritoneal shunt might be required [4, 5].

To diagnose and monitor post-haemorrhagic ventricular dilatation and to assess whether intervention is required, two-dimensional (2-D) cranial US is used. The most important measurements are the ventricular index, anterior horn width and thalamo-occipital distance [6]. These measurements aid in managing post-haemorrhagic ventricular dilatation, i.e. in determining when to perform lumbar punctures, whether to insert a ventricular reservoir, how often and how much cerebrospinal fluid to withdraw and whether to insert a ventriculo-peritoneal shunt. Another important prognostic imaging marker is the ventricular volume. This, however, cannot be measured using 2-D cranial US. Magnetic resonance imaging (MRI) is generally considered to be the most accurate way to determine the ventricular volume in neonates. Not all premature babies undergo MR scanning. When they do, it is usually performed once or twice during the neonatal period and at most centers the MRI is performed around term-equivalent age.

Although MRI provides high-resolution and anatomical detail, MRI has limitations. It can be challenging to maintain physiological stability during an MRI examination. Moreover, there are potential risks for infants during transfer to the MR unit, sedation might be required, and the examination is time-consuming and often costly. Also, MRI cannot be easily repeated for monitoring and follow-up. Two-dimensional cranial US is routinely used on the neonatal intensive care unit to diagnose and assess the progression of post-haemorrhagic ventricular dilatation. Therefore, it would be advantageous if US could also be used for ventricular volume assessment [7].

The purpose of the study was to assess whether the ventricular volume can be determined accurately using US, to pave the way for a US-derived measurement of ventricular volume. The rationale is that the ventricular volume might be more directly linked to intracranial pressure, which in turn is considered to be the actual cause of additional damage to the surrounding white matter.

In total, 31 preterm infants with germinal matrix intraventricular haemorrhage were eligible for inclusion in this study, and 14 infants (45%) developed post-haemorrhagic ventricular dilatation (Table 1). T1- and T2-weighted MR images were obtained at term-equivalent age (median 41.1 weeks, range 40.3–43.1 weeks). Ventricular volumes were obtained by automatic segmentation. The automatic segmentation was manually reviewed and corrected where necessary.

Within 3 days of the MRI examination, coronal and sagittal cerebral cranial US images were acquired (Fig. 1); at this time no ventricular taps from a Rickham reservoir were performed. For each child the ventricular index, anterior horn width and thalamo-occipital distance were measured on both sides and used for further statistical analysis as separate measurements (Table 2). We determined the intraclass correlation coefficient (ICC) using the one-way random effects models, yielding ICCs of 0.924 (95% confidence interval [CI] 0.854–0.961), 0.988 (95% CI 0.975–0.994) and 0.984 (95% CI 0.951–0.995) for the ventricular index, anterior horn width and thalamo-occipital distance, respectively.

Figure 3 shows the ventricular volume plotted as a function of ventricular index, anterior horn width and thalamo-occipital distance. We performed a linear fit to add a trendline to the plot, which resulted in the following relations for each of the covariates: ventricular volume (ml) = −16.1+1.77 × ventricular index (mm) (R² = 0.334); ventricular volume (mm) = −0.01+1.84 × anterior horn width (mm) (R² = 0.658); and ventricular volume (ml) = −3.70 + 0.576 × thalamo-occipital distance (mm) (R² = 0.574).

We fitted a general linear model to the data using forward variable selection (P<0.05) to obtain a model that can be used to predict the ventricular volume using 2-D cranial US measurements. For further analysis we chose to use a transformation with the cube root rather than a log transformation because this better explains the relationship between the physical quantities. That is, the model predicts volumes that scale as the third power of the linear measurements. Forward variable selection showed that the model with anterior horn width and thalamo-occipital distance had the highest adjusted R² and provided a fit with the most symmetrical distribution of residuals, resulting in a model that can be used to predict ventricular volume (∛ventricular volume [ml] = 1.096+0.094 × anterior horn width + 0.020 × thalamo-occipital distance [mm] with R² = 0.831). More details are given in Table 3. The residuals of our model are shown in Fig. 4.

The model as described here was validated against the model of Benavente-Fernandez et al. [13] (ventricular volume [ml] = −11.02+0.668 × ventricular index [mm] + 0.817 × anterior horn width [mm] + 0.256 × thalamo-occipital distance [mm]) [13]. Figure 5 shows the predicted ventricular volume as a function of the measured ventricular volume for both models and the residual error as a function of the measured ventricular volume for both models. The figure shows that larger discrepancies can be observed for ventricular volume greater than 20 ml, whereas smaller discrepancies of up to 5 ml were observed for ventricular volumes smaller than 20 ml.

In this study, we found that 2-D cranial US ventricular measurements of infants with germinal matrix intraventricular haemorrhage were significantly associated with the volume of the ventricles as determined with MR images in a wide range of ventricular volumes. The proposed model provides us with an estimate of the ventricular volume that can be obtained by performing only a few linear measurements on US images, which enables bedside estimation of ventricular volume. This highlights the utility of cranial US in the diagnosis and monitoring of preterm infants with post-haemorrhagic ventricular dilatation, and should encourage neonatologists to perform linear cranial US measurements of the cerebral ventricles. Figure 5 shows us that larger deviations from the model seem to appear for volumes over 20 ml. Although more data are needed to confirm this hypothesis, it seems that the model struggles to predict the larger volumes accurately. This might be explained by the fact that the shape of the ventricles changes considerably for larger volumes, so the estimation of the volume based on two local linear measurements might be more likely to fail. We argue that this does not necessarily hamper the use of the model in clinical practice. Infants with ventricular volumes higher than 20 ml undergo MRI regardless of the exact results of the US examination. Especially for smaller ventricular volumes, the model might aid in the discrimination between healthy infants and those with mild cases of germinal matrix intraventricular haemorrhage.

Other relevant work on the prediction of ventricular volume using US has been performed by Benavente-Fernandez et al. [13]. They created a model to predict the ventricular volume as measured by 3-D cranial US using linear 2-D cranial US measurements for infants with post-haemorrhagic ventricular dilatation. There were two major differences between this model (ventricular volume = a + b × ventricular index + c × anterior horn width + d × thalamo-occipital distance) and our model (∛ventricular volume = a + b × anterior horn width + c × thalamo-occipital distance) that might explain the observed discrepancies. First, the Benavente-Fernandez model included the ventricular index, anterior horn width and thalamo-occipital distance, whereas we only included the anterior horn width and thalamo-occipital distance. Second, their model did not perform a cube root transformation of the ventricular volume (∛ventricular volume). Other differences might be explained by the fact that 3rd and 4th ventricles were not included in the 3-D volume measurements by Benavente-Fernandez et al. [13], whereas they were included in our segmentations. The 3rd ventricle is hard to visualize on US imaging and, hence, few reference values are known in the literature. Moreover, the relative overestimation of ventricular volume on MRI as compared to 3-D cranial US has also been reported by Kishimoto et al. [14], which might be attributed to the fact that the posterior horns are difficult to visualize.

Limitations of this retrospective study include errors in the measurement of the variables apart from the variance in the data owing to the anatomical variation among patients and the intraobserver variability. For example, errors in the measurement of the ventricular index, anterior horn width and thalamo-occipital distance with 2-D cranial US might have occurred because of the absence of contrast between the borders of the ventricles and the rest of the surrounding brain tissue, especially for the occipital horns. Errors in the measurement of the ventricular volume on MRI might be caused by susceptibility artifacts or partial volume effects. Finally, errors in the model parameters could have been introduced by differences between the actual volume of the ventricles at the time of the MRI and 2-D cranial US acquisition; to minimize this difference, we chose a maximum of 3 days between the MRI and 2-D cranial US acquisition.

The reliability of our model is affected by the variance of linear 2-D cranial US measurements and because few data points are available for ventricles with a large volume. The fact that linear measurements with large variance are used, and that the distribution of several of the measured parameters is asymmetrical, could impact the prediction. This could also explain the fact that larger discrepancies between our model and the Benavente-Fernandez model were observed for larger volumes.

Although this study aimed to improve ventricular volume determination in clinical practice, the relation between ventricular volume and intracranial pressure — which might be the actual cause of damage — remains unclear, and additional investigations performed simultaneously, such as near infrared spectroscopy (NIRS), amplitude integrated electroencephalogram (EEG) and Doppler measurements of cerebral blood flow might provide additional important information [15, 16].

With the use of 2-D cranial US an approximation of the ventricular volume can be provided, which might be helpful for better management of post-haemorrhagic ventricular dilatation. Validation of our model in a larger study group including different time points and ventricular sizes is warranted.