I. Introduction
It is extremely crucial for the parents to know if their babies are
healthy or not during the pregnancy. The earlier the fetal abnormality
could be found, the more chance the abnormality could be cured.
Therefore, increasing attention and tremendous efforts have been put on
improving the harmlessness, effectiveness, and robustness of fetal
abnormality screening by researchers.
Although radiological examinations are known to be more precise on
revealing abnormalities of human body, the side-effects are also
significant. To avoid such potential, threaten to the fetus and the
mothers, a less risky technology is required for the screening. Thus,
ultrasound which has been proved to be less harmfulness, less expensive,
and more convenient than other radiological techniques is widely
utilized for the fetal abnormality screening. However, ultrasound
examination is less automatic than other radiological exams as the
sonographer needs manually hold the probe to conduct the examination.
Therefore, more experienced and skillful sonographers are required for
the ultrasound screening otherwise the results could not be reliable.
Unfortunately, the shortage of experienced sonographers has been a
severe problem ever since the born of ultrasound, and a qualified
sonographer may need years for training which is extremely time
consuming. To solve these problems, computer aided technologies are
urgently demanded.
In the screening, the priority for a sonographer is to find a series of
clinical standard planes which are a set of anatomical views of the
fetus. Then, the diagnosis is to be made based on both subjective
evaluation which is the observation of the sonographer, and objective
evaluation which includes several physical measurements. Obviously, the
final diagnosis is highly depending on the quality of standard
anatomical views acquired by the sonographer. In other words, as far as
the standard fetal anatomical planes can be precisely obtained, the
diagnosis is to be more accurate and more reliable.
For the retrospective studies of the fetal growth and diseases, well
organized and categorically arranged retrospective data are always
required. However, that could usually be not true since all of the
historical data are stored without any appropriate arrangement for most
of the hospitals. Therefore, a technique which can automatically
differentiate and separate the historical data according to the clinical
meaning or usage is to be beneficial to the researches.
In this work, a deep neural network based framework is presented for
classifying various types of standard anatomical planes of fetal head,
i.e., Transventricular plane (TV), Transthalamic plane (TT),
Transcerebellar plane (TC), Coronal view of eyes (Eyes), Coronal view of
nose (Nose), and other non-standard fetal ultrasound images
(Background). In the proposed framework, a YOLO based object detection
network is applied to locate the head region in ultrasound image, and a
set of powerful classification networks are utilized with model stacking
technique to give a final judgement on each image based on the detected
head regions. The contributions of this work are as follows: first of
all, this is the first piece of work using deep learning technology to
identify TV, TT, and TC to the best knowledge; secondly, this is the
first piece of work which successfully applied YOLO on this topic;
thirdly, the design of combining object detection network, object
classification network, and model stacking technique is novel to this
area; finally, the proposed framework achieves the state-of-the-art
performance.
The rest of this paper is arranged as follows. In Sec. II, related works
are briefly introduced. Then, the proposed framework is to be presented
in Sec. III. Experimental results and discussion are reported in Sec.
IV. The conclusion is finally made in Sec. V.