Page 234 - Computational Retinal Image Analysis

4 Lesion detection and segmentation 231

with shallow neural network for exudates (EX) classification. Grisan and Ruggeri
[32] presented a two-stage classifier for BL and RL detection. A Bayesian classifier
was used for pixel classification and followed by linear discriminative analysis that
performed lesion classification. Massey and Hunter [33] combined spatial clustering
of objects with multilayered neural networks and SVMs to detect BL and RL.

4.3 Region growing

Region growing methods examine neighborhoods of seed points and determine
whether they should be part of a specific region. Spencer et al. [34] proposed matched
filtering and bilinear top-hat transformation to find initial MAs. Subsequently, they
used a region growing algorithm to produce final segmentation results. Cree et al.
[35] extended [34] with more intensity-based features and novel region growing
algorithm. Frame et al. [36] combined region growing approach with three different
classifiers: shallow neural network, linear discriminant analysis, and rule-based
system. They concluded that manually created rule-based system provides best
results; however, their experimental dataset consisted of only 20 images. Fleming
et al. [37] emphasized the importance of contrast normalization and proposed
a “Watershed Retinal Region Growing” algorithm to segment lesions. Li and
Chutatape [38] proposed a fusion of edge detection and region growing algorithm for
exudate detection. Sinthanayothin et al. [39] introduced the “moat operator” that was
used to sharpen lesions’ edges. They combined it with the recursive region growing
algorithm to segment both BL and RL.

4.4 Thresholding
Threshold-based methods exploit differences in color intensity between various
image regions. Zhang et al. [40] proposed a method based on dynamic thresholding
and correlation coefficients of a multiscale Gaussian template to detect MAs. They
used 31 manually designed features based on intensity, shape, and response of a
Gaussian filter. Pereira et al. [41] combined a thresholding approach with the ant
colony optimizer to segment EXs. EX candidates were identified using a thresholding
method, whereas the unsupervised ant colony optimizer was used to enhance EXs’
edges. García et al. [42] suggested a combination of adaptive and global thresholding
approaches to find EX candidates. Next, a radial basis function classifier was
designed to classify lesions based on features mainly derived from lesions’ shape and
color. Saleh and Eswaran [43] developed a decision support system to find RLs. They
combined H-maxima transform for candidate selection with multilevel thresholding
to create final segmentation. Phillips et al. [44] used a combination of global, local,
and adaptive thresholding algorithms to find lesions. To tackle uneven illumination
of FPs, Sánchez et al. [45] combined dynamic thresholding based on mixture models
with edge detection.