Page 221 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 7 Early detection and diagnosis using deep learning 211

Recently, ML techniques that account for the intercorrelation
between regions have become pivotal component of computer-
assisted analytical methods and have been widely employed for
the automated diagnosis and analysis of neuropsychiatric disor-
ders. There is a long list of the different ML-based models that
have been used for the early detection of such neurological
disorders, but the DL-based model provides direction toward
the solution in a way like none other.
DL models make use of an end-to-end learning design philos-
ophy, which allow a system to use raw data as input, thereby
allowing them to automatically discover highly discriminative
features in the given training data set. This is the fundamental
basis of DL. The primary advantage of end-to-end learning is
the possible optimal performance that arises from all steps in
the processing pipeline being simultaneously optimized. There
are four levels of hierarchy, the fourth one being the maximum
(full). Another advantage of end-to-end learning is that it is an
effective visual explanation for why the classification decision
ability of CNNs is possible. The explanation helps the medical
practitioner to understand the behavior of the CNNs and to
discover new biomarkers. One of the best advantages of using
CNN is that it reduces the overall complexity by inserting convo-
lution and pooling layers, which in turn reduce the total number
of model parameters. CNN, when there is enough data available,
is widely used in the field of imaging and visual recognition
because of its effectiveness.
DNN, restricted and deep Boltzmann machine, deep belief
network, autoencoders as well as stacked and sparse autoen-
coders are DL methods that have been used for AD diagnostic
classification to date patients from cognitively normal (CN)
controls or MCI, which is the prodromal stage of AD. Each
approach is used to detect and work on how and when an AD
gets transformed into MCI with the aid of multimodal neuroi-
maging data.
The early detection of AD has been worked upon by scientists
all across the world for quite some time now. There have been
various methodologies developed such as representation
learning, softmax regression, and ROI sensitivity values. One of
the proposed models uses a deeply CNN, which carries out four
basic operationsdconvolutions, pooling, batch normalization,
and rectified linear unit. Each layer is connected to every other
layer using a dense connectivity pattern (Fig. 7.6).