Discussion
Main Findings
This study is the first to predict the fetal lateral ventricular width
using CNN-based DL algorithms. Our study shows that the scheme can
automatically pick out brain images from all stored freeze-frame images.
The sensitivity and specificity for brain images were 100% (376/376)
and 96.9% (376/388), respectively. The scheme can recognize TV and TT
planes and extract the brain regions. The sensitivity and specificity
for TV-TT planes were 97.6% (205/210) and 99.5% (205/206),
respectively. For the regression model, the MAE of the predicted lateral
ventricular width was 1.01 mm. More than 65% test images had a MAE of
less than 1 mm. If we used the 610 cases with lateral ventricular width
less than 15 mm to train and test the model, the MAE was 0.54 mm and
more than 82% test images had a MAE of less than 1 mm. The heat maps
provide evidence that our regression model predicting the lateral
ventricular width was based on the anatomical structure of lateral
ventricular.
Strengths and Limitations
Many psychiatric and neurodevelopmental disorders are associated with
enlargement of the lateral ventricles thought to have origins in
prenatal brain development [47]. Moreover, VM is one of the most
commonly detected fetal anomalies at the mid-trimester ultrasound (US)
and occurs in up to 2 per 1000 births [39-40]. Therefore,
recognizing this anomaly precisely and as early as possible is very
important.
A previous study [35] used deep learning algorithms to classify
fetal brain ultrasound images from standard axial planes as normal or
abnormal. However, it is not suitable to combine together the
ventriculomegaly cases and other CNS anomaly cases to train the
classification model. One reason is that the number of VM cases is much
higher than other CNS anomaly cases. In our dataset, from all the 22616
pregnant women, there are 90 VM cases (including 16 hydrocephalus cases)
and only 24 other CNS anomaly. Another reason is that, predicting VM
will focus only on the lateral ventricular region, while different other
regions may be evaluated for other CNS anomaly prediction.
Moreover, the study [35] limited the ultrasound images as standard
axial planes, while our study used all TV and TT planes. This is a real
problem that many cases have no standard axial planes stored and
inexperienced scanner may not be able to find out the desired standard
planes. The authors claimed that 70690 out of 92748 cases contained no
eligible standard axial neurosonographic planes and only about 16000
images can be used. In our study, the lateral ventricular width can be
measured in more than half of the TV and TT planes and we have 1431
available images from 626 cases.
In this study, we did not have any scale reference in the images and the
resolution of images were different. The ratio of the brain regions to
the whole images were also not the same. To solve this problem, we
detected and extracted the brain regions first and then resize the brain
regions into a same size. Experiment results shown that this was a
feasible way to mitigate the influence of these kinds of difference.
This study used only 626 pregnant women with gestational age between 22
to 26 weeks to train and test the modes. We got a MAE of 1.01 mm for the
first experiment, which use all the 626 cases to train and test the
regression model, and a MAE of 0.54 mm for the second experiment, which
use the 610 cases with lateral ventricular width less than 15 mm to
train and test the model. If we use more data, such as the data from the
third trimester of pregnancy, to train the models, the MAE would
potentially be reduced.
The lateral ventricular width is a continuous value. For
ventriculomegaly the threshold is 10 mm. For our models, we recommend a
smaller threshold like 9 mm or 8 mm. If the predicted lateral
ventricular width is bigger than this value, doctors should pay
attention to this fetal. A relatively small threshold can reduce false
negative prediction, which may lead to serious consequence. However,
false positive prediction is inevitable. In the first experiment, 53
images were predicted with lateral ventricular width bigger than 9 mm.
Among them, 30 were actually bigger than 10 mm and the ground truth of
other 23 images ranging from 8.3 mm to 0.99 mm. On the other hand, only
a small fraction of fetuses has large lateral ventricular width, if our
models can filter out most cases with small lateral ventricular width,
the workload of doctors can be reduced hugely.
Interpretation
Although the lateral ventricular width is usually measured in TV planes,
some doctors may measure this value in the TT plane or a transitionary
plane between TV and TT planes. In our dataset, a considerable portion
of TT planes were stored and used to measure the lateral ventricular
width. Furthermore, if the lateral ventricular width is very large, it
is usually hard to distinguish between TV and TT planes. For these
reasons, we used TV and TT planes for lateral ventricular width
estimation.
We built two models to predict the lateral ventricular width. The second
one was to use the 610 cases with lateral ventricular width less than 15
mm, to train and test the model. The reason is that, severe fetal
ventriculomegaly with lateral ventricular diameter >15 mm
(also sometimes classified as fetal hydrocephalus) is unusual and their
ultrasound images are much different from those of normal or mild fetal
ventriculomegaly cases. This kind of cases can be detected using
algorithms classifying fetal brain as normal or abnormal, such as study
[35] did. Furthermore, after ignored these cases, the performance of
the model improved remarkably.
After training the regression model to predict the lateral ventricular
width, we generated heat maps and their corresponding overlay images for
all test images. We found that, for the first experiment, all the heat
maps were activated at the left-upper corner. We guess the model used
the left-upper corner of each image to train something like a base value
to lower the overall MSE. The final predicted lateral ventricular width
combined the so-called base value with the value related to the lateral
ventricular region. If the lateral ventricular width was small, the
model might not detect the lateral ventricular region, and the final
predicted lateral ventricular width would be only determined by the
left-upper corner of the image. This was not very precision, but it was
safe for images with small lateral ventricular width. It was similar for
the second experiment that, the predicted lateral ventricular width of
most images with small lateral ventricular width were based on other
areas rather than the lateral ventricles.
It was worth noting that some images had markers on the lateral
ventricles. Was it possible that the models localized the lateral
ventricles and predicting lateral ventricular width using these markers?
From the heat maps we can see that, some images with large lateral
ventricular width and without markers were activated on the lateral
ventricles, while some images with small lateral ventricular width and
with markers were not activated on the lateral ventricles. We can
conclude that the regression model predicting lateral ventricular width
did not depend on the markers.