7th International Symposium on Image and Signal Processing and Analysis (ISPA 2011)
September 4-6, 2011, Dubrovnik, Croatia
Ensemble-based exudate detection in color fundus images Brigitta Nagy
Bálint Antal
[email protected] Faculty of Informatics, University of Debrecen Debrecen, Hungary
[email protected] Faculty of Informatics, University of Debrecen Debrecen, Hungary
Balázs Harangi
András Hajdu
[email protected] Faculty of Informatics, University of Debrecen Debrecen, Hungary
[email protected] Faculty of Informatics, University of Debrecen Debrecen, Hungary
Abstract—Diabetic retinopathy causes blindness to millions in the world. Exudates are early lesions of this disease so the automatic detection is very important to slow down the progression of retinopathy. In this paper, an ensemble-based system is proposed to improve the detection. Optimal combination of preprocessing methods and exudate candidate extractors are found and organized into a voting system for this aim. Our results show that in this way we outperform the individual exudate detector algorithms.
I.
INTRODUCTION
Diabetic retinopathy is a common cause of blindness especially in developed countries. However, at an early stage an appropriate treatment slow down the progression of this disease. Thus, the recognition of the early signs has great importance in corresponding automatic screening systems. Exudates are primary signs of diabetic retinopathy and occur when lipid or fat leaks from blood vessels or aneurysms. Exudates are light, small spots, which can have irregular shape, thus automatic exudate detection is a difficult task. We can find a large number of exudate detection algorithms in the medical image processing literature as see in [1]. These algorithms do not find all exudates and detect some false candidates, that is why we create a combination of these algorithms. The literature of ensemble-based systems suggests that the combination of individual algorithms built upon different principles tend to outperform individual accuracies. In this paper, we follow this principle with a complementary application of preprocessing methods. In other words, we will show how an optimal combination of pairs can be organized into a voting system. The evaluation of the proposed method in a publicly available database shows that this approach can outperform the application of individual algorithms. The rest of the paper is organized as follows: we list the considered preprocessing methods and candidate extractors in section 2 and 3, respectively. Section 4 exhibits a stochastic
Special Sessions Image and Signal Analysis for Computational Life Sciences
search algorithm to find the optimal ensemble, and section 5 gives the creation of the voting system. Comparative results for the ensemble and the individual detectors are included in section 6, and conclusions are drawn in section 7. II.
PREPROCESSING METHODS
In this section, we present the preprocessing methods selected as possible participants of the ensemble. These preprocessing methods are applied on the images before performing the candidate extractors in order to enhance image contrast regarding exudates. The results of the preprocessing methods can be seen in Figure 1. •
Illumination Equalization (IE) [2]: Pixels are adjusted in the following way: I eq (r , c) = I ( r , c) + m − I w ( r , c )
,
where I(r,c) is the original intensity value, m is the desired average intensity and in a window. •
I w(r,c) is the mean intensity value
Gray-World Normalization (GN) [2]: r new =
r g b , g new = , and b new = RAvg GAvg BAvg
,
where (r, g, b) are the original intensity values, RAvg, GAvg and BAvg are the average intensity values in each band and (rnew, gnew, bnew) are the new values, respectively. •
Division by an over-smoothed (DS) [3]: The mean of the intensity values is computed within a window. The original intensity value is divided by the mean value of its neighborhood.
700
7th International Symposium on Image and Signal Processing and Analysis (ISPA 2011)
•
Morphological contrast enhancement (MC) [3]: Morphological opening and closing are applied. T Then the new intensity values are subtracted from the original intensity values multiplied by three.
•
Background subtraction (BS) [3]: The background is estimated by computing the mean of the intensity values within a w1 × w1 window. The he mean of the previously
September 4-6, 2011, Dubrovnik, Croatia
•
( [9]: In this paper, Fuzzy C-Meanss clustering (FCM) Sopharak et al. proposed a method using fuzzy c-means c clustering in order to determine whether a pixel is exudate or non-exudate. They used four features to obtain a coarse segmentation result. Then, morphological operations are applied to refine the segmentation result.
•
Naive Bayes (NB) [10]: Sopharak et al. proposed an exudate extractor algorithm which used a pixel-based classification. Five features based on a classic machine learning approach are extracted to classify pixels.
•
Split-and-merge (SMF, SMC) [11]: The exudate detection was divided into two steps in this method. In the coarse step (SMC), the high local variances between the exudates and background are found. The algorithm splits the images into disjoint regions and then merges them in the fine step (SMF) if they were enough homogenous. A histogram-based based thresholding was used for the merged parts as adaptive thresholding.
•
Hard-Exudates (HE) [12]: This method uses pixel-based classification, ification, but the training database is extracted from the currently analyzed image. First, the algorithm de detects the sure exudates and uses them as positive pixels to set the training database. database Finally, the rest of the pixels are classified by their related relat properties.
calculated values are computed in a w2 × w2 window are subtracted from the intensity value of each pixel. •
Vessel removal and extrapolation (VE) E) [4]: The complete vessel system is removed. The inpainting algorithm [5] [ is used to fill the holes.
(a) IE
(b) GN
(c) DS
IV. (d) MC
(e) BS
(f) VE
Figure 1. Results of preprocessing reprocessing methods methods.
III. CANDIDATE EXTRACTOR ALGORITHMS In this section, we present the selected exudate candidate extractors. Three morphological operation and four classification based methods are selected as possible participants for the ensemble. Binary images are required as the output of the candidate extractors for the consequent voting scheme. scheme •
Morphological based (Mo) [6]: Walter et al. proposed a method using morphological operations such as morphological closing in order to eliminate blood vessels. Standard deviation is calculated, and thresholding threshold is applied to find the candidate regions.. Next, morphological reconstruction is applied with a second thresholding to find the contours of the exudates.
•
Optimal morphology (Om) [7]: Sopharak et al. proposed a method based on optimally adjusted morphological operations. The optic disc is removed, and Otsu’s algorithm is used to threshold hreshold during the detection process.
•
Coarse-to-fine (CTF) [8]: This technique is based on morphological operations and H--maxima transform. Contrast enhancement on L channel is applied after the image is converted from RGB to LUV color space.
Special Sessions Image and Signal Analysis for Computational Life Sciences
ENSEMBLE CREATION
We create an ensemble pool by creating all possible pairs of preprocessing methods and exudate candidate extractors. extractor In the case of pair, the given preprocessing method is applied on the image before performing the given candidate extractor algorithm. A simulated annealing-based based search algorithm [14] [ is used to find the optimal ensemble of pairs. Each ach combination is evaluated using the following energy function: E = − FScore = −
2 * sensitivit y * PPV , sensitivit y + PPV
(1)
which is expected to be minimized by the optimal combination. In the evaluation of E, sensitivity and PPV stand, stand for:
sensitivity =
TP , TP + FN
(2) TP positive predictive value ( PPV ) = TP + FP .
701
7th International Symposium on Image and Signal Processing and Analysis (ISPA 2011)
September 4-6, 2011, Dubrovnik, Croatia
TABLE I.
PREPROCESSING METHOD, CANDIDATE EXTRACTOR PAIRS SELECTED AS OPTIMAL COMBINATION.
Next, we give the detailed description of this algorithm. ALGORITHM I. FINDING THE OPTIMAL ENSEMBLE OF PAIRS .
Input: • An initial
temperature
M ∈ R> 0 and
a
minimal
temperature
M min ∈ R >0 . • A temperature change q ∈ R with 0 ≤ q ≤ 1 . • A set H = { PPi , CE j | i = 1,..., K , j = 1,..., L}
containing
all
pairs, where K is the number of preprocessing methods and L is the number of candidate extractors, respectively. • A search space SS = P(H ) , where P is the power set containing all subsets of pairs. • A function random(V), which chooses a random element of the set V. • A function accept : ℜ × ℜ × ℜ> 0 × [0,1] → {true, false}, which is defined in the following way: e − e' true, exp > r, accept(e, e' , M , r ) = M false, otherwise.
• An energy function E : SS → ℜ , as described before. Output: • s optimal
∈ SS , where E ( s optimal ) = min E ( s) . s∈SS
Algorithm: 1. s←random(SS) 2. e←E(s) 3. SS←SS-{s} 4. while SS≠0 and M>Mmin do 5. s’←random(SS) 6. e’←E(s’) 7. SS←SS-{s’} 8. if e’