11 Segmentation and Visualization of Drusen …
313
11.3.2.3 GA Candidate Region Extraction
The coarsely segmented GA regions obtained in the previous iterative step, are still
insufficient to be considered as an initialization outline for the CVLSF method.
Isolated low-intensity false negative regions within correctly detected GA regions
and extensive false positive regions (as can be observed in Fig. 11.20e) tend to
cause “leakage” (segmentation expansion to neighboring structures) in Chan Vese
methods, yielding sub-optimal results. A GA candidate region extraction refinement
is considered here with the goal of further including isolated background regions and
excluding false positive locations in the CVLSF model initialization outline.
11.3.2.4 Segmentation of GA Regions Based in an Improved C-V
Model via Local Similarity Factor
The results obtained after the coarse segmentation refinement are taken as an initialization for an improved region-based C-V model [67] with a local similarity factor
(CVLSF), which is introduced here to suppress noise influence, while guaranteeing
detail preservation in the segmentation results. The objective function for partitioning an image I (x, y) ∈ into two regions (GA region and background) is defined
as:
E(c 1 , c 2 , C) λ 1
¨
in(C)
|I (x, y) − c 1 |
2 + L SF 1 (x, y)
dxdy
+ λ 2
¨
out(C)
|I (x, y) − c 2 |
2 + L SF 2 (x, y)
dxdy + μLength(C)
(11.5)
where μ ≥ 0 is fixed constant parameter, and λ 1 > 0, λ 2 > 0 control the contributions of the internal energy and external energy terms, respectively, where object
regions taken as internal term are the inside of the contour C (in(C)) and background
regions considered as external term are the outside of C (out(C)). Using the level set
definition [70] to represent C, that is, C is the zero level set of a level set function
(x, y), we can rewrite this objective function as:
E(c 1 , c 2 , ,(x, y)) λ 1
¨
|I (x, y) − c 1 |
2 H ((x, y))
+L S F 1 (x, y)H ((x, y)))dxdy
+ λ 2
¨
|I (x, y) − c 2 |
2
(1 − H ((x, y)))
+L S F 2 (x, y)(1 − H ((x, y))))dxdy
+ μ
¨
δ((x, y))|∇(x, y)|dxdy
(11.6)
313
11.3.2.3 GA Candidate Region Extraction
The coarsely segmented GA regions obtained in the previous iterative step, are still
insufficient to be considered as an initialization outline for the CVLSF method.
Isolated low-intensity false negative regions within correctly detected GA regions
and extensive false positive regions (as can be observed in Fig. 11.20e) tend to
cause “leakage” (segmentation expansion to neighboring structures) in Chan Vese
methods, yielding sub-optimal results. A GA candidate region extraction refinement
is considered here with the goal of further including isolated background regions and
excluding false positive locations in the CVLSF model initialization outline.
11.3.2.4 Segmentation of GA Regions Based in an Improved C-V
Model via Local Similarity Factor
The results obtained after the coarse segmentation refinement are taken as an initialization for an improved region-based C-V model [67] with a local similarity factor
(CVLSF), which is introduced here to suppress noise influence, while guaranteeing
detail preservation in the segmentation results. The objective function for partitioning an image I (x, y) ∈ into two regions (GA region and background) is defined
as:
E(c 1 , c 2 , C) λ 1
¨
in(C)
|I (x, y) − c 1 |
2 + L SF 1 (x, y)
dxdy
+ λ 2
¨
out(C)
|I (x, y) − c 2 |
2 + L SF 2 (x, y)
dxdy + μLength(C)
(11.5)
where μ ≥ 0 is fixed constant parameter, and λ 1 > 0, λ 2 > 0 control the contributions of the internal energy and external energy terms, respectively, where object
regions taken as internal term are the inside of the contour C (in(C)) and background
regions considered as external term are the outside of C (out(C)). Using the level set
definition [70] to represent C, that is, C is the zero level set of a level set function
(x, y), we can rewrite this objective function as:
E(c 1 , c 2 , ,(x, y)) λ 1
¨
|I (x, y) − c 1 |
2 H ((x, y))
+L S F 1 (x, y)H ((x, y)))dxdy
+ λ 2
¨
|I (x, y) − c 2 |
2
(1 − H ((x, y)))
+L S F 2 (x, y)(1 − H ((x, y))))dxdy
+ μ
¨
δ((x, y))|∇(x, y)|dxdy
(11.6)