Page 206 - Computational Retinal Image Analysis

202 CHAPTER 11 Structure-preserving guided retinal image filtering

also proposed in [10] for optic cup segmentation.
In recent years, deep learning approaches have been applied for optic cup
segmentation as well. In the M-Net introduced above [12], the cup is segmented
jointly with the optic disc. In [29], the joint disc and cup segmentation is combined
with image-level features for direct glaucoma detection. In [40], fast R-CNN is
adopted to segment the optic cup. Besides the deep learning approaches, we have
also proposed a direct approach to compute the cup diameters using sparse learning
and its variations [11, 41, 42].

Joint optic disc and optic cup segmentation
Because of the relative location constraint between the optic disc and optic cup,
the optic disc boundary could provide some useful prior information for optic cup
segmentation, for example, shape constraint and structure constraint [43]. The work
in [10, 25] deals with disc and cup in two separate stages with different features, where
the cup is segmented from the disc region. Zheng et al. [44] integrated the disc and
cup segmentation within a graph-cut framework. However, they consider the disc and
cup as two mutually exclusive labels, which means that any pixel can only be assigned
to one label (i.e., background, disc, or cup). Moreover, the method only employs
color features within a Gaussian mixture model to decide a posterior probability of
the pixel, which makes it unsuitable for fundus images with low contrast. In [30], a
modified U-Net deep network is introduced to segment the disc and cup. However,
it still separates disc and cup segmentation in a sequential way. In [45], an ensemble
learning method is proposed to extract disc and cup based on the CNN architecture.
An entropy sampling technique is used to select informative points, and then a graph-
cut algorithm is employed to obtain the final segmentation result. However, this
multiple-step deep system limits effectiveness in the training phase. We proposed a
M-Net [12] to segment the optic disc and cup simultaneously. Instead of segmenting
the cup from the disc region or segmenting the cup from the full image, the M-Net
adopts a multilabel loss function to achieve the joint segmentation of the optic disc
and cup. Fig. 2 illustrates the M-Net framework. M-Net is an end-to-end multilabel
deep network which consists of four main parts. The first is a multiscale layer used
to construct an image pyramid input and achieve multilevel receptive field fusion.
The second is a U-shape convolutional network, which is employed as the main body
structure to learn a rich hierarchical representation. The third part is a side-output
layer that works on the output of early convolutional layers to support deep layer
supervision. Finally, a multilabel loss function is proposed to achieve simultaneous
segmentation of OD and OC.

Image quality
A common challenge in analysis is the image quality [46], which has been neglected
in most earlier algorithms. Very often, low-quality images lead to poor performance.
There are many factors that might affect the image quality, including the cooperation
of the patient, the experience of the operator, the imaging device, the image