Fig 1.
A comparative study on image-wise enhancement techniques.
(A) Two example images from the DRIVE dataset. (B) The green channel of (A). (C) and (D) show the results of applying Histogram Equalization and Gamma correction image enhancement methods on (B), respectively. Each method enhanced image contrast: however, there still exist large areas of inhomogeneity. (E) Results after applying Retinex on (B). Retinex enhances the contrast between vessels and background well, and in consequence the vessels are more easily identifiable.
Fig 2.
Illustration of the enhancement effect on a test image using the local phase filter at different scales from 1 to 4.
Fig 3.
Enhancement results produced by an eigenvalue-based method [28], a wavelet-based method [17] and the local phase method, respectively.
An image was randomly chosen from each of the four datasets. From top to bottom: DRIVE, STARE, ARIA, and VAMPIRE. (A) Example images. (B) Eigenvalue-based enhancement results. (C) Wavelet-based enhancement results. (D) Local phase based enhancement results.
Fig 4.
Enhancement results on selected region with vascular bifurcation and crossover produced by an eigenvalue-based method [28], a wavelet-based method [17] and the local phase method, respectively.
(A) A randomly chosen image from the DRIVE dataset. (B) Selected region with vascular bifurcation and crossover. (C) Eigenvalue-based enhancement results. (D) Wavelet-based enhancement results. (E) Local phase based enhancement results.
Fig 5.
Relative importance of Retinex-based inhomogeneities correction.
(A) A randomly chosen image from the DRIVE dataset and expert’s annotation. (B) Vesselness map using local phase filter (top), and the segmentation result (bottom) when the Retinex is applied. (C) Vesselness map using local phase filter (top), and the segmentation result (bottom) when the Retinex is not used.
Table 1.
Segmentation performance of Retinex pre-processing algorithm with and without applied on segmentation framework.
Se: sensitivity; Sp: specificity; Acc: accuracy; AUC: area under the curve.
Fig 6.
Illustrative enhancement results using different methods and their subsequent graph cut based segmentation results.
(A) A randomly chosen image from the DRIVE dataset and expert’s annotation. (B)-(D) Eigenvalue-based (FR), wavelet-based (WL), and proposed local phase-based (LP) enhancements on (A). (E) Expert’s annotation. (F)-(G) Graph cut based segmentation results on (B)-(D).
Fig 7.
Overview of the main steps of our method and the comparison results obtained with other two segmentation methods.
(A) A randomly chosen image from the DRIVE dataset. (B) The green channel of (A): this channel has the highest contrast between regions of vessel and the background. (C) Results after applying Retinex on (B). Retinex successfully enhances the contrast between vessels and background, and the vessels are more easily identifiable. (D) Local phase map of (C): the edges of the vessels are enhanced, and made more visible, to make the vessel stand out further from the background. (E) Expert’s annotation. (F)-(H): Segmentation results with the level set (LS), Total variation (TV), and graph cut (GC) based segmentation methods, respectively.
Table 2.
Segmentation performance of different possible combinations of three enhancement methods (LP, WL, FR) and three segmentation methods (GC, TV, LS) on the DRIVE dataset.
LP, WL and FR denote local phase, wavelet and Frangi’s eigenvalue based enhancement filters respectively. GC, TV and LS denote graph cut, total variation and level set based segmentation methods respectively. Se: sensitivity; Sp: specificity; Acc: accuracy; AUC: area under the curve.
Table 3.
Segmentation performance of different possible combinations of three enhancement methods (LP, WL, FR) and three segmentation methods (GC, TV, LS) on the STARE dataset.
LP, WL and FR denote local phase, wavelet and Frangi’s eigenvalue based enhancement filters respectively. GC, TV and LS denote graph cut, total variation and level set based segmentation methods respectively. Se: sensitivity; Sp: specificity; Acc: accuracy; AUC: area under the curve.
Table 4.
Segmentation performance of different possible combinations of three enhancement methods (LP, WL, FR) and three segmentation methods (GC, TV, LS) on the ARIA dataset.
LP, WL and FR denote local phase, wavelet and Frangi’s eigenvalue based enhancement filters respectively. GC, TV and LS denote graph cut, total variation and level set based segmentation methods respectively. Se: sensitivity; Sp: specificity; Acc: accuracy; AUC: area under the curve.
Table 5.
Segmentation performance of different possible combinations of three enhancement methods (LP, WL, FR) and three segmentation methods (GC, TV, LS) on the VAMPIRE dataset.
LP, WL and FR denote local phase, wavelet and Frangi’s eigenvalue based enhancement filters respectively. GC, TV and LS denote graph cut, total variation and level set based segmentation methods respectively. Se: sensitivity; Sp: specificity; Acc: accuracy; AUC: area under the curve.
Table 6.
Performance of different segmentation methods, in terms of sensitivity (Se), specificity (Sp), accuracy (Acc) area under the curve (AUC), on the DRIVE and STARE datasets.
Fig 8.
Illustrative vessel abnormality detection result based on the proposed segmentation method.
(A) Original fluorescence angiography image. (B) Abnormality detection result. Red color indicates abnormal vessels and green color shows normal vessels.