Page 282 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 10 Deep neural network in medical image processing 273

of implementations of various deep learning techniques for diag-
nosis of ophthalmic diagnosis using various retinal image data
sets. Tajbakhsh et al. [24] reviewed various research papers
dealing with application of deep learning in biomedical image
processing with imperfect data sets.
Hesamian et al. [25] and Taghanaki et al. [40] presented a
critical review of popular methods by employing deep learning
techniques for medical image processing. A very interesting and
critical review of usage of deep learning techniques for detection
of Alzheimer’s disease was presented by Ebrahimighahnavieh
et al. [26]. Shen et al. [27] reviewed various deep learningebased
approaches of image registration, cellular structure detection,
and tissue segmentation, leading to computer-aided disease diag-
nosis and prognosis.
Another interesting two-step deep learning approach for anal-
ysis of whole-slide histopathologic images to detect lung cancer
lymph node metastases was given by Pham et al. [28]. Yang
et al. [29] presented an innovative method of predicting perfor-
mance of deep learning features by using CNN-F (CNN Fast)
based on staging CT images of sentinel lymph node. Dong et al.
[30] built a deep learning radiomic nomogram to predict the num-
ber of metastasizing lymph nodes in locally advanced gastric
cancer.
Li et al. [31] evaluated various deep learningebased algo-
rithms in clinical situations using more than 13,000 fundus
images from more than 9500 patients of diabetic retinopathy.
Torre et al. [32] proposed a deep learning interpretable classifier,
which uses a new pixel-wise score propagation model. Kamnit-
sas et al. [33] proposed a multilayer deep three-dimensional
CNN architecture, which employs a dense training scheme to
cut the computational load of processing 3D medical images.
Long et al. [34] proposed to adapt contemporary classification
networks such as GoogleNet [15]and AlexNet[6] to fully convo-
lutional networks and also defined a fine shallow layer to
produce detailed and accurate segmentations. Another impor-
tant original work in the field of detecting mediastinal lymph
node malignancy in computed tomography (CT) images is that
the network has been customized with an advanced activation
function leaky ReLU (rectified linear unit).
There has been a huge growth in the interest generated for
deep learning frameworks as work is picking up in the field of
image segmentation and other fields such as legal document
classification, medical image analysis, and video segmentation,
among others. There are full-fledged frameworks now being
introduced for deep learning. Torch [35] is one such framework,