Page 62 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

50 Chapter 2 Deep convolutional neural network in medical image processing

endoscopic testing for various diseases related to the abdominal
region, more research can be expected in this domain in the com-
ing years.

5. Critical discussion: inferences for future
work and limitations

A wide range of applications of deep CNN are very much
evident from the literature summary done in this chapter. There
are different CNN architectures studied that are applied to
various medical imaging modalities and different tasks associated
with clinical image processing. Different architectures studied
include cascading networks, deep networks, multiple-view net-
works, multiscale networks, and so on. It is observed that in
many works, limited data are used, and also expert annotations
are also rare. This architecture aims to decrease parameter space,
reduce the computational time as well as handle 3D images. In
several studies, it is observed that large data are in 3D format,
mainly the data collected from MRI and CT scans. These 3D
data are generally handled by converting the 3D image into 2D
slices or by combining different features from multiview and
2D planes so as to take advantage of contextual information. In
the present scenario, a huge amount of medical data is available,
which can be used by the researchers to develop more efficient
models for the betterment of the healthcare unit. Also, it is
seen that for a particular body part, only a few of the imaging mo-
dalities have been considered. This gives future scope for the re-
searchers to apply CNN to other modalities for the particular
anatomical structure.
Even though DL and particularly deep CNN gives better per-
formance in the field of medical imaging, there are few limita-
tions to look into. CNN models require a large amount of
computational power for its execution. Limited computational
power will require more amount of time to train the model that
will be subjected to the training data size. The CNN architecture
needs labeled information for the implementation of supervised
learning and manual labeling for the images, which is tough
work. These problems are gradually overcome as there is large
computational power available, large data storage capacity, and
also efficient deep CNNs.
Fig. 2.6A presents the distribution of articles in years related to
CNN application in healthcare and medical imaging. The data
used in the figure are collected from the Scopus database [120]
through searching the words “CNN in healthcare,”“convolution