Page 111 - Computational Retinal Image Analysis

104 CHAPTER 6 Retinal vascular analysis: Segmentation, tracing, and beyond

Other techniques also exist. For example, by establishing an analogy between
quantum mechanics and image processing, it is proposed in Ref. [70] to transform
image pixels to quantum systems such that they can be evolved from an initial state
to a final state governed by the Schrodinger equation.

3.2 Supervised segmentation
The alternative to the unsupervised paradigms are learning-based methods, in which
a set of training examples is provided to learn a model that is expected to segment
input retinal images at test time as well as the performance it has gained during
training. One early work is that of Akita and Kuga [11], where neural networks are
used in segmenting retinal vasculature. A system is developed in Ref. [71], where a
patch-based neural network model is learned by backpropagation to classify each
pixel as being vessel or not, then OD and fovea regions are obtained via template
matching.
A typical supervised approach is to first construct or develop a set of dedicated
features or filters, then to build a statistical model based on the features as sufficient
statistics, with model parameters being estimated (i.e., learned) from a set of training
examples. For example, a local patch-based approach is considered in Ref. [72], where
an AdaBoost classifier is in place to work with 41 features extracted from the local image
patch of the current pixel, where the pixel is to be predicted as being either vessel or not.
Martin et al. [73] devise gray-level and moment invariants-based features for segmenting
the retinal vessels using neural networks. Ricci et al. [74] work with orthogonal line
operators and support vector machine to perform pixel-wise segmentation. Becker et al.
[75] present a discriminative method to learn convolutional features using gradient
boosting regression technology. In Ref. [76], a pool of difference-of-Gaussian (DoG)
filters is used that after training, filters are adaptively selected to best fit the current
vessel of interest. A learning-based DoGs filtering approach is proposed in Ref. [77],
with one application focus being about the detection of vascular junction points, where
orientation is achieved via shifting operations. Empirically it is shown robust to contrast
variations and the presence of noises. Furthermore, many learning-based methods [78–
81] also advocate the automation of the feature learning process. For example, Soares
et al. [78] elaborate upon 18-dimensional Gabor response features to train two Gaussian
mixture models (GMMs), which are further employed to produce a binary probability
map for a test image. The method of Becker et al. [81] employs a gradient boosting
framework to optimize filters and often produces impressive performance.
Several learning paradigms, including graphical models and ensemble learning, are
presented with very promising results. A discriminatively trained, fully connected CRF
approach is developed by Orlando and Blaschko [82]. This is followed by Orlando [83]
with more expressive features in their fully connected CRF model. Besides, the work of
Fraz et al. [30] showcases an ensemble classification approach consisting of bagged and
boosted decision trees, with features from a variety of aspects including orientations of
local gradients, morphological transformations, and Gabor filter responses.