Page 281 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

272 Chapter 10 Deep neural network in medical image processing

Kazuyoshi Itoh [8] trained a neural network to identify 12 digits
(10 Arabic and 2 Roman) in an X-ray film using 22 sheets of film as
training data. Similar work in generic object recognition was also
done by Zuo Bai et al. [9] using powerful CNNs. In this chapter,
they proposed an extreme learning machine based on local recep-
tive field. CNNs were used very effectively by Joly et al.[10] for
identifying plants using a huge data set of dated and geotagged
plant photos. Lo et al. [3] used CNN for pattern recognition in
general medical image, which worked in conjunction with
enhanced training methods.
Murat et al. [11] published a comprehensive review of various
applications of deep learning techniques for ECG signal analysis.
Hao et al. [12] published a detailed survey on application of
deep learning algorithms for semantic segmentation; the paper
also discusses the various challenges involving the field. Another
full-blown use of deep learning techniques in the field of detecting
myocardial infarction and arrhythmia is discussed by Gopika et al.
[13]. An important turning point came in the field of image
processing when Navab et al. [14] proposed U-Net, a deep convo-
lutional networkebased framework, which made better use of
annotated transmitted light microscopy images. A highly effective
22-layer deep CNN named GoogLeNet was presented in Large-
Scale Visual Recognition Challenge 2014 by Szegedy et al. [15].
LeCun et al. [16] review the usage of various deep learning models
in the field of processing image, video, and speech; they also high-
light the benefit of using backpropagation algorithm.
Zhao et al. [17] implemented an innovative brain tumor seg-
mentation method by successfully integrating full-fledged CNNs
and conditional random fields, which produced segmentation
results with spatial and appearance consistency. One more inter-
esting research work was carried out by Zhang et al. [18]; they pro-
pose an innovative method of boundary detection by combining
machine learning methods and texture suppression in a unified
framework. Havaei et al. [19] propose a novel method of identi-
fying low- and high-grade glioblastomas from MRI images using
an approach, which uses both local and global features in a
CNN. A dual-staged deep learning framework was proposed by
Bria et al. [20] to train model for detecting small lesion in medical
images, which counters the problem of high class imbalance. Saba
et al. [21] discuss various deep learningebased algorithms to
analyze medical images as well as point out future growth areas
in the field.
Liu et al. [22] discuss the potential of various representative
deep learning models for lung cancer localization and subsequent
diagnosis. Sengupta et al. [23] presented a comprehensive review