Page 49 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 2 Deep convolutional neural network in medical image processing 37

the “shift-and-stitch” model [57]. In the proposed method, the
shifted versions of the input image are applied with fully CNN.
One can achieve the full resolution of the output with a loss of
pixels due to the valid convolutions by stitching the result together.
Authors in Ref. [58] extended the idea of the fully CNN to pro-
pose the U-Net network. The U-Net consists of a regular fully
CNN, which is followed by an unsampling part. In the unsampling
part, the upconvolutions are taken into consideration so as to in-
crease the size of the image, which are termed as expansive and
contractive paths. Long et al. [57] combined the aforementioned
network with so-called skip connections so that it can directly
connect opposing the expanding and contracting convolutional
layers. A quite similar idea was taken by Cicek et al. [59] for 3D
data. Milletari et al. [60] gave the concept of extending the
U-Net layout by incorporating ResNet-like residual blocks and a
dice loss layer instead of cross-entropy, which straightaway
reduces the commonly used measure for segmentation error.

4. Application of deep convolutional neural
network in medical image analysis

In this section, an overview of CNN's contributions to the
various anatomical areas of the human system is discussed.
Table 2.2 presents recent reviews on CNN application in medical
imaging related to different anatomical regions of the human body.

4.1 Brain
CNNs have been widely used for image processing of the brain
in various application areas. Multiple studies have discussed the
classification of Alzheimer's disease and the segmentation of
anatomical structures and brain tissues. Other important areas
in which various works have been done are segmentation and
detection of lesions. Most of the methods that are proposed
work in 2D, even though brain images are 3D volumes. This is
done by slicing the brain image and then examining slice-by-slice
of the 3D volumes. This is generally driven by either the thick sli-
ces compared with resolution of in-plane in some data sets or
reduced computational requirements. 3D networks have also
been employed in a few of the recent publications. Almost all
the work done related to brain image analysis has used the MRI
scans for the study. It can be expected that other image modal-
ities can also be benefitted by using DL methodologies. Various
studies have been summarized in Table 2.2.