Fig 1.
An example of color fundus image containing MAs.
(A) The color fundus image. (B) The corresponding enlarged part of (A) in green channel with indicated MAs (diamond: regular MA, circle: subtle MA, triangle: irregular MA, square: clustered MAs, pentagon: MA close to vessel). Reprinted from [8] under a CC BY license, with permission from Dr. Yalin Zheng, original copyright 2012.
Fig 2.
Schema of the proposed method.
The solid line path summarizes the training process of MAs detection, and the dashed line path summarizes the test process.
Fig 3.
Different distributions of gradient vectors of the vessel-like and the MA-like objects.
(A) The gradient field of a vessel-like object. (B) The distribution of gradient vectors in (A). (C) The gradient field of a MA-like object. (D) The distribution of gradient vectors in (C).
Fig 4.
The process of map computation.
(A) The preprocessed input image. (B)-(C) The log condition number maps with different support regions. (D) The final map. (E) An image patch of preprocessed image contains 5 true MAs marked with ‘□’. (F) The corresponding patch of
map.
Fig 5.
The process of vessel removal.
(A) The image morphologically reconstructed by using a binary map. (B) The vessel removed image by subtracting (A) from Fig 4A. (C) The enlarged part of marked region in (B).
Fig 6.
Some results of vessel removal in special cases.
(A) The image patch containing subtle MAs with different size. (B) The corresponding result after vessel removed. (C) The image patch containing a MA near vessel. (D) The corresponding result after vessel removed.
Fig 7.
The illustration of at different position with different direction θ.
(A)-(B) The different positions of the same MA. (C)-(D) The positions centered at one subtle MA and one non-MA. (E)-(H) are the polar coordinate plots of the direction θ versus the value of in the direction θ at these positions.
Fig 8.
The result of candidate MAs localization.
(A) The map. (B) The final location of candidates MAs, where all candidates indicated with ‘×’ and the true MAs provided by medical expert marked with ‘□’.
Fig 9.
Segmentation results of candidate MAs.
(A), (C) and (E) The image patches containing MAs. (B), (D) and (F) The segmentation results.
Table 1.
Comparison of different candidate extractors on ROC training set.
Table 2.
The statistical results of final candidate sets extracted from training set of different databases.
Fig 10.
The sensitivities of cross validation on ROC training set with different T in RUSBoost classifier.
Fig 11.
FROC curves produced by different classifiers.
Table 3.
Quantitative results of the ROC competition for each participating team.
Fig 12.
The FROC curves for each of different approaches on the test set of the ROC competition.
Fig 13.
FROC curves of the proposed method and the DRSCREEN approach on the DRDB database.
The FROC curve reproduced from the original work of DRSCREEN [25] on the same database.
Table 4.
The results of the proposed method and the DRSCREEN approach on the DRDB database.
Table 5.
Classification results of different feature combinations on ROC training set.
Fig 14.
(A)-(B) The true MAs with low contrast. (C) The vessel crossing with high contrast. (D)-(E) The non-MAs have very similar appearance as true MAs.