Page 294 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 10 Deep neural network in medical image processing 285

languages such as Cþþ, R, Python, and so on. Mxnet is
portable and scalable to use multiple GPUs.
• CAFFE [36], which is an acronym for convolutional architec-
ture for fast feature, is an open-source deep learning frame-
work developed by the University of California, Berkeley. It
is originally written in Cþþ and has a Python interface. It is
very popular with image classification and image segmenta-
tion developers.
• Tensorflow [38] is an extremely popular free-to-use open-
source deep learning framework. It was developed by Google
Brain Team for internal consumption but was later released
under BSD License in 2015.

4. Segmentation techniques in image
processing

Studies such as Nicholas et al. [41] are useful in evaluating the
field of accuracy evaluation of image segmentation but often
concentrate on the geometry of the objects examined and ignore
the possibility of adopting a controlled yet nongeometric
approach (e.g., Wang et al. 2004 [42]).Inaddition,supervised
methods are typically compared in a specific study case without
further and important issues being discussed, such as the suit-
ability of methods as a function of context. As image segmenta-
tion is increasingly being used in a wide range of applications,
it is assumed that perhaps the behavior and usefulness of spe-
cific methods may differ in each case. The selection of a method
for determining the preciseness of the image segmentation can
therefore be based on insufficient and potentially problematic
understanding of the options available. The segmentation
algorithm must be chosen from a range of available choices,
but comparative studies (e.g., Basaeed et al. 2016 [44]; Neubert
et al. 2008 [43]) are uncommon. Each single segmentation algo-
rithm, depending on the parameter setting, is also usually able
to produce an extensive number of outputs. Therefore, it is diffi-
cult to pick the most suitable segmentation. Fig. 10.5 represents
the types of segmentation techniques.

4.1 Different approaches for segmentation

Supervised approaches for image segmentation accuracy
analysis use reference data to measure the accuracy of the con-
structed objects. Reference data also consist of polygons derived
from remotely sensed information in use (e.g., based on visual