Page 205 - Computational Retinal Image Analysis

1 Introduction 201

a modified Chan-Vese model using texture features. In [27], edge detection and the
circular Hough transform are combined with an active shape model to extract the disc.
To overcome the limitations of pixel classification-based methods and deformable
model-based methods, we propose a superpixel classification-based method [10] and
combine it with the deformable model-based methods. With the rapid development of
convolutional neural network (CNN) in image and video processing [28], automatic
feature learning algorithms using deep learning have emerged as feasible approaches
for retinal image analysis. Recently, some OD segmentation algorithms [12, 29, 30]
based on deep learning, especially fully convolution network structure [31], have
been proposed. Sevastopolsky [30] adopted a modified U-Net to directly segment
the OD and optic cup for further CDR calculation and glaucoma diagnosis. In a
further extension of the U-Net, Fu [12] proposed a novel architecture called M-Net to
jointly segment OD and OC as well as a disc-aware method [29] for direct glaucoma
detection, where the OD segmentation is the of a series of operations. These methods
can greatly improve the performance of OD segmentation, based on the strong
learning capacity of deep learning.

Optic cup segmentation

Optic cup segmentation is another important task in retinal image analysis. Detecting
the cup boundary from 2D fundus images without depth information is a challenging
task as depth is the primary indicator for the cup boundary. In 2D fundus images, one
landmark to determine the cup region is the pallor, defined as the area of maximum
color contrast inside the disc [23]. Another landmark is the vessel bends at the
boundary of the cup [26, 32]. Compared with disc segmentation, fewer methods
have been proposed for cup segmentation from 2D fundus images. Thresholding is
used to determine the cup in [33–35], relying on intensity difference between cup
and neuroretinal rim. A level set-based approach is used in [36]. It relies on the edges
between cup and neuroretinal rim. This method and thresholding-based methods are
essentially based on pallor information. However, in many subjects from screening
data, there is no obvious pallor or edge within the disc to mark the cup boundary. In
[37], small vessel bends (“kinks”) near the initially estimated cup have been used to
aid the cup segmentation. The challenge is to exclude vessel bends from a boundary
not belonging to the cup, especially when the initial estimation is inaccurate. A similar
concept is used in [25] to locate relevant-vessel bends (“r-bend”) near a pallor region
determined by bright pixels. This method, again, requires the pallor information to
determine a good initial estimation of the cup boundary. Moreover, it requires at least
a few bends in the nasal, inferior and superior angular regions of the disc for the cup
boundary fitting, which is not necessarily true for many images in our experience.
Xu et al. [38] proposed a sliding window and regression-based method. Although this
performs better than earlier methods, the sliding window strategy imposes a heavy
computational cost. Yin et al. [39] developed a deformable model-based method for
cup segmentation, where the initialization of the cup boundary is based on pallor
combined with prior knowledge of the cup. Previously, superpixel classification was