Interval type-2 fuzzy approach for retinopathy detection in fundus images

Abstract

In the manuscript, an automatic approach for analysis and detection of various stages of retinopathy defects in human eyes has been proposed. The approach consists of a robust preprocessing technique of the retina fundus image to mitigate the effects of noise and poor lightening in the image. To realize a compressive analysis of the defects, methods for extracting blood vessels and optic disc in the fundus image has also been developed. Adaptive Histogram Equalization (AHE), median filtering and Connected Component Analysis techniques were used in separating blood vessels and optic disc from each fundus image. The pre-processing utilizes canny edge detection and Morphological Closing on the fundus image. An interval type-2 fuzzy (IT2F) clustering is applied to segments an input image into four clusters. These four clusters from the fuzzy segmentation are further analyzed to extract various stages of retinopathy abnormalities (e.g., Hemorrhage, hard exudates etc.). The extracted blood vessels and optic disc are removed from the analysis to enhance the defects detection process. Experiments were conducted on DIARETDB1 database. The experimental results obtained are validated using the ground-truth images contained in DIARETDB1 database. Impressive results are recorded throughout the experiment. Hard-Exudates and Hemorrhage were detected from the fundus images and results from similarity indexes such as, accuracy (94.11%) sensitivity (93.03%) and specificity (98.45%) were recorded.

Keywords

Retinopathy fuzzy c-means connected component analysis canny edge detection morphological operation high pass filter (HPF)

1. Introduction

In medical imaging, research on automatic diagnosis of eye defects (diabetic retinopathy) from digital fundus images has been active for a very long time. The vast interest vested by researchers in this field is largely due to the prevalence medical needs and urge for new products in the medical industry that will significantly reduce the health care costs and improve diagnosis accuracy. Diabetic retinopathy, sometimes referred to as diabetic eye diseases, presents a serious threat to human sights and potential loss of vision. The process of detection of such abnormalities in the retina is known as diabetic retinopathy and has become quite significant to be carried out at its early stage of manifestation. To detect retinopathy, segmentation, and identification of key tissues such as blood vessels and eye disc are of great importance.

Segmentation of blood vessels are mostly divided into two techniques, namely: rule-based and pattern recognition-based techniques [1]. According to [2, 3], rule-based techniques mainly include matched-filtering technique, tracking of vessel, morphological operations, multi-scale methods and model-based methods. Techniques based on Pattern Recognition (PR) are subdivided into two namely: supervised and unsupervised techniques. Both methods require clustering and classification processes. In supervised approach a labeled training data with ground-truth images are required and after clustering each pixel can be identified and grouped into pixel vessels or non-pixel vessels. On the contrary of the case of supervised method, in the unsupervised method, no previous labelled training data are needed. Both methods (supervised and unsupervised) involve clustering techniques, that divides the data into groups of clusters [4].

The labeled training and testing data are mainly the basis of supervised techniques. The method categorizes pixels into two groups as vessel pixel or non-vessel by deploying the training data of the labeled sample images. Supervised techniques are identified with higher precision scores for well labeled and annotated dataset and are seriously affected for a poorly curated dataset. On the other hand, supervised techniques are considerably susceptible to false edges [4].

Roy Chowdhury et al. [5] proposed a three-level retinal blood vessel segmentation algorithm. The first, a threshold binary image is obtained using HPF and the other binary threshold image is gotten by top-hat reconstruction of the red fields in a green plane image. Thereafter, the common region from both binary images is extracted and merged. The classification process is obtained from the image after merging. All pixels that are classified as vessels are merged with dominant vessels to obtain the vascular structure. Furthermore, improving the image is done using post-processing methods. The major vessels and the fine vessels are accurately segmented using the suggested technique. It has low dependency on the trained-data and contain less calculations.

Hassanien et al. [2, 3] proposed fuzzy cluster fitness function, together with artificial optimization of bee colony technique to identify and extract the thickness of blood vessels. The approach utilizes two different stages in the formation of clusters. For the first clustering, optimization techniques based on artificial bee algorithm which was implemented with the aim of extracting blood vessels. The authors used Fuzzy C-Means (FCM) and the objective function. The next phase significantly improves the location of cluster centers by utilizing local search to extract blood vessels with a thinner and well as small in diameter.

Emary et al. [4], applied possibilistic FCM clustering technique on the classical FCM objective function. For an improved result, a search technique based on cuckoo search was applied. This search method was used purposely to optimized possibilistic FCM clustering method. Finally, results were obtained by implementing the approach on a DRIVE dataset. It was claimed to have been capable of detecting abnormalities such as hemorrhages, exudates, and changes in epithelium pigment.

Franklin et al. [6] suggested a technique based on blood vessels recognition in retinal images by using a multi-layer perceptron neural network. The method derived input from the basis three-color channels. Despite some success recorded their approaches, the problem is far from being solved due to some drawbacks such as low performances of the proposed approaches, complexity in implementation and overall lack of comprehensiveness.

Authors in [7] a proposed a method for Diabetes Classification using type-2 fuzzy inference systems with better parameter designs based on genetic algorithm. The optimized the parameter values of Type-2 Trapezoidal membership functions and the type of model (Mamdani or Sugeno) with genetic optimization techniques. Results using their techniques was compared with type-1 fuzzy systems on PIMA Indian Diabetes. An improved performance was reported.

There are several variants of fuzzy clustering algorithms such fuzzy c-means (FCM), possibilistic c-means (PCM) and fuzzy possibilistic c-means FPCM have been reported in the literature and used in many applications requiring clustering and classification such as Diabetic retinopathy [4, 8]. However, for better handling of uncertainty within the dataset an improvement of such techniques is required [9]. The is achieved through application of Type-2 Fuzzy Logic techniques to baselines (FCM and PCM), which produces Interval Type-2 Fuzzy C-means (IT2FCM) and Interval Type-2 Possibilistic C-Means (IT2PFCM) [8]. In recent times extension of other baseline clustering algorithms (PCM and FPCM) using Type-2 Fuzzy Logic techniques were also proposed [8]. Similarly, researchers on Diabetic retinopathy are also emerging using deep convolutional neural network (DCNN) learning. Authors in [10], proposed a method Diabetic Retinopathy Classification based on convolutional neural networks and Hierarchical Genetic Optimization. The used several pre-processing with a hierarchical genetic algorithm (HGA) with the pre-processing method to obtain higher classification accuracy using CNN model.

In this manuscript, an automatic process for extraction and detection of various stages (e.g., hard exudate and [2] Micro aneurysms and hemorrhage) of eye diabetic retinopathy has been proposed. The proposed techniques implemented a variant fuzzy-type 2 clustering (Interval Type-2 fuzzy IT2FCM) for anomaly extraction in fundus image. Prior to IT2FCM clustering an effective technique for blood vessels and optic disc segmentation was as so proposed. The method includes extractions of the blood vessels and optic disc from the fundus image. Using results from interval type-2 fuzzy (IT2FCM) clustering, various categories of retinopathy were analyzed and detected. The major contributions of the are as follows:

Proposed an efficient and effective techniques for blood vessels and optic disc segmentation via adaptation of computationally efficient, existing image processing techniques.

Proposed a method for exudates region extraction and eventual detection of retinopathy using background subtraction and adaptation of IT2FCM.

The article is organized into sections as follows: Section 1 contains introduction with review of literature, Section 2 presents some theoretical background and motivation of the algorithms used. In Section 3, proposed approach is presented while Sections 4 and 5 contain experimental results and discussion and Section 6 presents the conclusion.

2. Theoretical background and motivation

The motivation of this research is to performs both feature extraction (extraction of blood vessels, optical disc, exudates etc.) and classification of the retinopathy diseases using computationally efficient image processing techniques that would produce an acceptable and competitive performance levels with other available techniques. For instance, the choice of IT2FCM over other variants is due to its capacity to handle uncertainty and noise in the data with less computation. Under this section several algorithms relevant to the proposed approach are discussed in summary.

2.1 Fuzzy c-means (FCM)

Fuzzy clustering techniques is a kind of soft decision classification approach whereby a data point is allowed the flexibility of belonging to more than one classes with different degrees of membership to each data class or cluster. It is essentially a kind of unsupervised learning algorithm and one of the well-known variants of this algorithm is fuzzy c-means (FCM) clustering [11].

Consider a dataset $X$ with $n$ data points belonging to a data space with an $s$ dimension as:

$X = {x_{1}, x_{n}, \dots x_{n}} \subseteq R^{s}$ (1)

The objective of the fuzzy c-means is partitioning the data points into $k$ number of clusters by minimizing an objective function $J (U, C)$ , where $U$ is the membership matrix and $C$ is the clusters center vector:

$J (U, C) = \sum_{i = 1}^{n} \sum_{k = 1}^{c} {(u_{i k})}^{m} x_{i} - c_{k}^{2}$ (2)

Whereby the term $x_{i} - C_{k}^{2}$ is the euclidean distance of data point $x_{i}$ from the centers $C_{k}$ of the k clusters. This distance measure could be replaced by any convenient distance metric operator [11].

The membership function matrix $U (u_{i k})$ is given by Eq. (3)

$u_{i k} = {(\sum_{j = 1}^{k} {(x_{i} - C_{k})}^{2 / (m - 1)})}^{- 1}$ (3)

Where $m$ is a fuzzifier coefficient and always $m >$ 2. The necessary condition for the membership function matrix is such that the sum of membership of each data point $x_{i}$ in all the clusters equals to unity as shown in Eq. (4).

$\sum_{i = 1}^{c} u_{i k} = 1 0 ⩽ u_{i k} ⩽ 1$ (4)

The center of each cluster is also updated at each iteration as given by Eq. (5).

$C_{k} = \frac{\sum_{i}^{n} {(u_{i k})}^{m} x_{i}}{\sum_{i}^{n} {(u_{i k})}^{m}}$ (5)

The membership matrix and the clusters centers are continuously updated as the objective function is minimized through the iterative means.

2.2 Interval Type-2 FCM

Interval Type-2 Fuzzy (IT2FCM) is basically an extension of FCM clustering by applying Type-2 Fuzzy techniques. IT2FCM is capable of modelling uncertainties using what is referred as Footprint of Uncertainty (FOU), where such uncertainties cannot be modeled by type-1 fuzzy sets. It uses two fuzziness parameters $m_{1}$ , $m_{2}$ to make FOU, corresponding to upper and lower values of fuzzy clustering. With these fuzzifiers two different objective functions $J_{m_{1}}$ (upper) and $J_{m_{2}}$ (lower) are to be minimized as follows [12].

${\begin{matrix} J_{m_{1}} (U, c) = \sum_{i = 1}^{N} \sum_{k = 1}^{C} {(u_{i k})}^{m_{1}} d_{i k}^{2} \\ J_{m_{2}} (U, c) = \sum_{i = 1}^{N} \sum_{k = 1}^{C} {(u_{i k})}^{m_{2}} d_{i k}^{2} \end{matrix}$ (6)

Where $d_{i k} = x_{i} - c_{k}$ is the Euclidean distance between a data point $x_{i}$ and a cluster center $c_{k}$ , $N$ is the number of data points (patterns) and $C$ is the number of data clusters (centroid). The membership function $u_{i k}$ have both upper $\bar{u_{i k}}$ and lower $\underline{u_{i k}}$ forms which are determined as follows [12].

$\bar{u_{i k}} = {\begin{matrix} \frac{1}{\sum_{j = 1}^{C} {(d_{i k} / d_{j k})}^{2 / (m_{1} - 1)}} if \frac{1}{\sum_{j = 1}^{C} (d_{i k} / d_{j k})} < \frac{1}{c} \\ \frac{1}{\sum_{j = 1}^{C} {(d_{i k} / d_{j k})}^{2 / (m_{2} - 1)}} if \frac{1}{\sum_{j = 1}^{C} (d_{i k} / d_{j k})} ⩾ \frac{1}{c} \end{matrix}$ (7) $\underline{u_{i k}} = {\begin{matrix} \frac{1}{\sum_{j = 1}^{C} {(d_{i k} / d_{j k})}^{2 / (m_{1} - 1)}} if \frac{1}{\sum_{j = 1}^{C} (d_{i k} / d_{j k})} > \frac{1}{c} \\ \frac{1}{\sum_{j = 1}^{C} {(d_{i k} / d_{j k})}^{2 / (m_{2} - 1)}} if \frac{1}{\sum_{j = 1}^{C} (d_{i k} / d_{j k})} ⩽ \frac{1}{c} \end{matrix}$ (8)

As a result of each pattern or data point has two membership functions $\bar{u}$ (upper) and $\underline{u}$ (lower), two centers are computed for each cluster [12]. The two centers, upper or right center $\bar{c}$ and lower or left centers $\underline{c}$ , are both computed using Eqs (5) and (6). To obtain a single centroid $C_{i}$ for a particular cluster, type reduction is applied to the two clusters above as follows [12].

$C_{i} = (\bar{c_{i}} + \underline{c_{i}}) / 2$ (9)

2.3 Connected component analysis

The Connected Component Analysis (CCA) is an image processing technique that has useful applications in many fields. It is applicable to different format of image being grey, colored, or binary [13, 14]. The algorithm begins with a scan for the entirety of the image on a pixel-by-pixel operation from the top leftist pixel of an image to the bottom pixel on the furthest right corner of an image. In the process all, neighboring pixels found to have shared same designated intensities are marked as a single connected object. Each marked connected pixel possess a unique label in the output image. For instance, in binary images, only pixels with intensity value equals to logical ‘1’ are designated as an object whereas those with logical ‘0’ as intensity values are marked as non-object pixels. For grayscale images range of intensity value is used (i.e., 0–50, 60–80) to designate an object pixel if its intensity falls within the predetermined range otherwise is labelled as non-object. A slight alteration to the described approach is used when CCA operates in a color-space images (i.e., RGB). The algorithm operates on basically two different level of connectivity windows within the neighborhood of the center pixel (i.e., 4 and 8 connectivity). Figure 1 shows an example of CCA algorithm on a binary image with 4-neighborhood connectivity to extract Binary Large Objects (BLOBs)

Figure 1.

4-Neighborhood connectivity operator.

For every marked object during the operation, it is treated as a unique BLOB and with its labeled values and corresponding indices. For each object in the CCA vector, numerous geometrical features are computed. These features include the area, perimeter, centroid, eccentricity and so on for each object which vital for different analysis.

2.4 Morphological operation

Morphological operation, in the context of computer vision, refers to the operation carried out in two dimensional (2D) spatial domains of an image to describe the properties, shapes and area of objects on the image plane. The operation becomes useful in the processing task of a set of points to reveal vital features of an object. The set of points in the image spatial domain represents an object. Four major morphological operators include dilation, erosion, opening and closing. While dilation increases the size of the image, erosion thins the image. Morphological opening and closing are compound operators which allow to fill an inner hole or help in getting rid of a small fragment on the image, respectively [13, 14]. The two major inputs to the morphological operators are the input image and, a specially created image called, Structural Element (SE). The choice of the SE is arbitrarily and mostly depending on the type and the nature of the operation intended. In most instances, SE is much smaller than the processed image since it is a form of area description of sets of binary points (object) within the image plane [13, 14].

2.5 Canny edge detection

Canny Edge Detection operators is one of the most prominent and effective edge detection operators proposed in [15]. The operator uses five basic steps in detecting edges in image: (1) Preprocessing where Gaussian filter is used to smoothen and remove noise from the image, (2) Compute the intensity gradients of the image, (3) Non-maximum suppression approach which is to remove spurious response to edge detection, (4) Double threshold is use to determine potential edges in the image and (5) Detect edges using Hysteresis by dropping all edges that are weak and have no connection to strong edges [15, 16, 17, 18].

Figure 2.

Flowchart of the proposed algorithm.

3. Proposed methods

There are basically three fundamental parts in the proposed approach which includes (a) extraction of features (b) retinopathy defects detection (c) classification of stages and evaluation. The feature extraction encompasses the segmentation of blood vessels, optic disc. They provide a comprehensive result, so the subsequences analysis is made easier. The proposed approach starts with the input RGB fundus image and extracting green channels. This followed with features extraction to segment regions of blood vessel and optic disc. Th segmented blood vessels and optic disc were utilized with background subtraction techniques and IT2FCM clusters to identify a stage in retinopathy. The second stage of the proposed approach applies a sequence of preprocessing (median filtering, gamma-transformation, canny edge detection and morphological operation) to enhance the image quality. IT2FCM is then applied on the enhanced image to group it into four clusters. The choice of the number of clusters 4 is based on the experimental trials with different clusters numbers like 2, 6 and 8 but the results were not as good as with 4. The final stage is exudate detection. Figure 2 provides the overview of the proposed method and the intermediary results.

3.1 Features extraction

The feature extraction includes blood vessel segmentation and optic disc segmentation and finally the IT2FCM for getting best clusters. The entire process was carried out on a public database DIARETDB1. The DIARETDB1 database consists of 89 color fundus images. Out of these images 84 of them consist of a minimum of one mild non-proliferative signs of the diabetic retinopathy. In accordance with the experts involved in the evaluation process, 5 of the images are considered normal without any signs of the diabetic retinopathy [19].

3.2 Blood vessel segmentation

Blood vessels segmentation is one of the fundamental aspects of automatic retinal image analysis. Even though some attributes of the blood vessels like branching patterns and vessels width play important role in diagnosis, moreover it’s also useful that when removed from the retinal image, it enables better analysis of retinal defects that occurs in a non-blood vessels portion.

To extract the blood vessels, series of image processing techniques were applied. Firstly, the RGB colored fundus image is separated into three channels (R, G and B). Green channel is used due to its higher information contents compared to others. A median filter is applied to the G-channel to smoothen it. This operation is being followed with Adaptive Histogram Equalization (AHE) [20]. AHE ensures that the contrast within the retinal fundus images are adjusted and suppresses some of the inherent uneven lightning during the acquisition of the image as shown in Fig. 3d.

Figure 3.

Pre-processing. (a) original fundus image (b) Green-Channel of Fundus image (c) Histogram of G-Channel (d) AHE of G-channel (e) Histogram of (d).

The next stage is to estimate the background information including the optic disc from the contrast-adjusted image (Fig. 3d). Morphological opening operation is applied to the compliment version (negative) of the input image (Fig. 3d) using a structural element of ‘ball’ type since the aim is to initially remove the optic disc region and then the background from the input image. The output of this operation is subtracted from the inverse of image (Fig. 3d) to give a mask image with optic disc removed. Thereafter, background information is obtained by firstly applying median filter to the mask image to further suppress noise. secondly morphological opening used with a structural element of ‘disk’ shape and a bigger radius than in the previous operation. Since the output of these two operations estimates the background information of the image, the background image is removed from the output of the median filter image which leaves behind an image whose foreground contains the blood vessels and some noises. This estimate is adjusted and eventually converted to binary image (Fig. 4a). This estimate of the blood vessels come with noise which can easily be removed. The method used here to remove this noise is using CCA. Initially all the BLOB objects areas within the image are identified and then any BLOB whose area is less than 2% of the larges BLOB within the image is revived. This operation gives the noise-free blood vessels in Fig. 4b.

Figure 4.

Blood vessels. (a) noisy blood vessels (b) noise-free blood vessels.

3.3 Optic disc segmentation

The optic disc represents the optic nerve head, and it is the brightest region in the retinal fundus image where blood vessels converge. Locating Optic disc becomes paramount because it provides useful information in identification of defects in the fundus image. To achieve that, the estimate of the optic disc determined during the blood vessels segmentation is used. It is obtained by applying morphological operation to the compliment of image in (Fig. 3d) using a structural element of ‘ball’ type (Fig. 5a). The contrast of this image is adjusted which results in the image in Fig. 5b. The optic disc region becomes much distinct in the low intensity regions, a hard decision thresholding is applied to remove all pixels of higher intensity magnitude greater than 50. The hard thresholding results in the output image in Fig. 5c. The final stage is to remove some noise using CCA in a similar way to blood vessels extraction. The only difference here is that areas and centroids of each BLOB object is estimated and two area thresholds (upper $=$ 40,000 and lower $=$ 15,000) are used to remove the noise. The last remaining object is retuned as the optic disc and its centroid is used to draw an ellipse around its boundaries as in Fig. 5d.

Figure 5.

Optic disc. (a) optic disc estimate (b) intensity adjusted optic disc estimate (c) noisy optic disc (d) extracted optic disc with drawn ellipse.

3.4 IT2-FCM clustering

In the clustering stage a further preprocessing is applied. Median filtering is applied to the green channel (Fig. 3b) to reduce the effect of noise. To mitigate the effect of poor contrast in the image. Power Law (Gamma) transformation which is a pixel-wise operation is used to transform the intensity of a particular pixel based on the Power Law (Gamma) transformation equation in Eq. (10).

$s = c r^{γ}$ (10)

Where $r$ is the input pixel $I (i, j)$ , $s$ is the new transformed value of $r$ and $c$ and $γ$ are positive constants. Canny edge [15] detection is applied to the output of the gamma operation is give a rough estimate of the edges of the vital objects within the image. The canny edge detection comes at cost the image is left behind is hollow and need to be refilled to re-introduce some of the vital information removed by the operation. The output of this operation is fed into the Type-2 Fuzzy (IT2FCM) clustering process as described in chapter two. IT2FCM returns four clusters are formed are subsequently used for the detection and classification stage as shown in Fig. 6.

3.5 Detection and classification

It is interesting to note that the four clusters have all the important features associated to the retinopathy diseases and they all share correlations among themselves. For example, the cluster number 4 contains all the blood vessels in the original fundus image. Similarly, cluster 4 is exact compliment of the summation of the remaining 3 clusters. Therefore, most of the detection and classification of the various stages in retinopathy are drawn from these clusters.

Figure 6.

Clusters segmentation using Interval Type-2 fuzzy (IT2FCM).

3.6 Exudates detection using background subtraction

These are lipid parts of the retina defects which are yellowish and bright. To extract Exudates from the cluster results, all the features extracted from a particular image are considered (i.e., Optic disc, Blood Vessels and IT2FCM clusters). For exudate detection experimental results from IT2FCM reveals that always all the information related to region containing exudates are found in Cluster 1 group (Fig. 7c). Hence only Cluster 1 group are considered from the 4 groups extracted using IT2FCM. It is also noted that additional information found in the Cluster 1 group are traces of either Optic Disc or Blood vessels. Therefore, Exudates detection is now reduced to removal of any trace of blood vessels or optic disc from the cluster 1 group (Fig. 7a–d). This is realized by subtracting both the blood vessel (Fig. 7a) and optic disc (Fig. 7b) from the cluster using logical AND gate operation. The operation resulted in the detection of exudates in Fig. 7d. It is also worth noting that the blooded vessels are dilated a bit before being used to ensure effective removal of any trace of vessels in the cluster.

Figure 7.

Exudates detection. (a) extracted optic disc (b) extracted blood vessels (c) IT2FCM cluster 1 result (b) detected exudate using background subtraction.

4. Simulation results

The entire process was carried out on a public database DIARETDB1. The DIARETDB1 database consists of 89 color fundus images. Out of these images 84 of them consist of a minimum of one mild non-proliferative signs of the diabetic retinopathy. In accordance with the experts involved in the evaluation process, 5 of the images are considered normal without any signs of the diabetic retinopathy [19].

The images were taken in the Kuopio university hospital. The images were selected by the medical experts, but their distribution does not correspond to any typical population. During the experiments all sample fundus images were used in accordance with the proposed procedures to finally obtain four clusters for each sample fundus image. The evaluation of the results is based on the stages of the diabetic retinopathy developments (e.g., micro-aneurysms, hemorrhage, exudates, etc.). Since the ground-truth results are supplied in the database [19], these ground-truth images are used to compute the accuracy of the proposed method.

4.1 Evaluation

To evaluate our findings of the proposed approach, some standard evaluations procedures were adopted. Initial six different similarity indexes were used to compare the detected abnormalities with the ground-truth images in the database. The similarity measures deployed includes Sensitivity, SpecifityAccuracy [19], Dice index [21], Jaccard index [22] and Structural Similarity (SSIM) index [23] and Positive Prediction Value (PPV) [24]. The indexes were computed both at image level and the overall database. On the other hands a similar comparison was made between the results obtained using the proposed method and other state-of-the-arts algorithms in the literature with similar work. The original ground-truth images in DIARETDB1 database supplied come with annotation marks made by the medical expert on them. These annotations are firstly removed before comparison is drawn. Figure 8 presents a sample of annotated ground-truth image containing hard-exudate defects.

Figure 8.

Annotated ground-truth hard exudates image.

In Table 1 experimental constants used during the experiments are tabulated where Table 2 present experimental results of exudates detection in some of the images from the database (in Fig. 9). In Table 3, we provided comparison between the proposed method and the results obtained by other authors in the literature.

Table 1

Experimental constants

Technique	Parameter name	Optimum value or range
Canney edge	Canney lowl imit	0.08
Canney edge	Canney upper limit	0.18
Power law gamma	Gamma value	0.7
Second order laplacian filter	Size	[2 2]
Unsharp mask	constant	0.8
Interval type-2 fuzzy	Fuzziness m1,m2	(2.5, 1.5)
Connected component analysis	Threshold max, min	40000, 15000

Table 2

Image level experimental results for hard exudates

Image	Specificity (%)	Sensitivity (%)	Dice (%)	Jaccard (%)	SSIM (%)	PPV(%)	Accuracy (%)
Image01	99.20	86.78	90.23	95.12	99.89	99.99	99.19
Image02	92.44	85.01	97.82	98.91	99.00	99.98	92.43
Image03	98.56	88.76	91.47	95.74	99.81	99.99	98.54
Image04	86.67	82.16	88.77	94.39	98.29	99.79	86.63
Image05	93.99	98.14	82.63	91.31	99.23	99.98	94.02
Image06	95.95	95.61	83.46	91.73	99.48	99.98	95.95
Image07	95.97	84.07	94.48	97.24	99.48	98.98	95.95
Image08	98.84	99.83	89.43	94.72	99.85	99.99	98.84
Image09	93.59	97.40	78.64	89.32	99.18	99.97	93.62
Image10	99.33	99.19	88.69	94.35	99.91	99.99	99.33

Figure 9.

Samples fundus images from the database whose results are presented. The first row is the original fundus images, second row is the groundtruth responding images results, the third row is our target and the final row is our responding results images.

4.2 Discussion

The Retinopathy diabetes is a very serious problem which begin to manifest itself as a tiny blood dot in the retina blood vessels (microaneurysms) which mostly go unnoticed. Prolonged microaneurysms degenerate into hemorrhages and eventually exudates. These are very vital stages which when not early and properly attended to may leads to serious degeneration to other stages which difficult to treat and may result in total or partial blindness.

It is quite a big challenge detecting early stages of these process. This is because at early stages their textures resemble that of the other parts of the retina like macula and blood vessels. To be able to detect any of that at early stage, a very robust algorithm is needed which can make distinction between different textures and features of the fundus image despite their similarities.

In order to achieve that, the proposed approach has taken many scenarios into account and provide a comprehensive approach to segmentation of major features in the fundus image (e.g., optical disk, blood vessels) to reduce the effects of false detection in the classification process. The approach utilized a much robust feature extraction process (using Interval Type-2 FCM) which has performed significantly well

Table 3
Comparisons with different methods

Authors	Method	Accuracy of classifincation	Sensitivity	Specificity	Positive predictive value (ppv)
Muhammad Faisal et al. 2014 [25]	Support Vector Machınes (SVM)	Not reported	96.9%	100%	100%
R. Radha et al. 2013 [26]	Gmorphologıcal Process and Clusterıng Technıque	98%	Not reported	Not reported	Not reported
Sumandeep Kaur et al. 2017 [27]	K-Means Colour Compression and Fuzzy Logic	96%	94.7%	Not reported	Not reported
Akara Sopharak et al. 2009 [28]	Using Fuzzy C-means Clustering	87.28%	99.24%	42.77%	24.26
Acharya et al. 2009 [29]	Blood vessel, exudates, microaneurysms, haemorrhages	86%	82%	86%	Not reported
Vujosevic et al. 2019 [30]	Single lesions	Not reported	82%	92%	Not reported
R.H.N.G. Ranamuka et al. 2020 [31]	Fuzzy Logic	Not reported	75.43%	99.99%,	Not reported
Pavle et al. 2016 [32]	Deep neural networks and anatomicallandmark	Not reported	78%	Not reported	78%
T. Walter et al. 2002 [33]	Means of morphological reconstruction techniques	92.8%	Not reported	Not reported	92.4%
E. Imani et al. 2016 [34]	Signal separation algorithm	89.01%	99.93%	82.64%	Not reported
Abdullah Saeed et al. 2016 (Saeed, Alharthi, and Emamian, 2016)	Digital analysis and mathematical morphology operations	86%	80%	84.69%	Not reported
A.M. Ashir et al. 2021 [24]	Diabetic Retinopathy Detection Using Local Extrema QuantizedHaralick Features with Long Short-Term Memory Network	95.45%	91.65%	95.45%	99.34%
Cordero-Martínez Rodrigo and Sánchez D, 2022 [35]	CNN	94.0%	Not reported	Not reported	Not reported
Abiyev R, Altiparmak H, (2021) [36]	T2FNN (32 rules)	87.0%	85.0%	89.0%	Not reported
Abiyev R, Altiparmak H, (2021) [36]	T2FNN (82 rules)	99.1%	98.7%	99.3%	Not reported
Abiyev R, Altiparmak H, (2021) [36]	T2FNN (100 rules)	96.9%	98.4%	96.1%	Not reported
Proposed Method	Interval Type-2 FCM	99.1%	99.0%	99.2%	99.59%

compared to most of the similar literary works we compared with in the Table 3. The process of separately segmentation of Optical disk and blood vessels have been very challenging aspect of the retinopathy diabetes detection which has been handle with a lot of success by the proposed algorithm.

5. Conclusion

A method for feature extraction, detection and evaluation of retinal diabetic has been proposed. A comprehensive and through analysis of the retinal images using Interval Type-2 fuzzy techniques with segmentation capability of vitals features such as blood vessels, optic disc has been presented. To realize the desired performance, the proposed approach has taken many instances into account by designing effective approach to segmentation of major features in the fundus image (e.g., optical disk, blood vessels) to reduce the effects of false detection in the classification process. The approach utilized a much robust, better noise and uncertainty handling clustering method IT2FCM to obtain competitive results which significantly works well. The process of separately segmentation of Optical disk and blood vessels have been very challenging aspect of the retinopathy diabetes detection which has been handled with a lot of success by the careful design and use of existing image processing techniques. Other effective state-of-the-art methods using variants of fuzzy type-2 techniques and deep convolutional neural networks can perform excellently well and even better but comes at the cost of increase in computation and memory usage. Future works can deploy such techniques with optimization of computation and memory usage.

The results obtained from the process are consistent across all the metrices used to evaluate the algorithm and it can clearly be seen that a lot of improvement has been recorded over the other methods that have mainly used Fuzzy Logic and other segmentation algorithms. The detection and identification method for defects has also been developed with higher efficiency. The results are validated using the ground-truth image and the standard performance matrices which is the gold-standard for evaluating new algorithms. The results are impressive and competitive with relevant state-of-the-arts research as shown from the experiments.

References

Zarandi

MHF

Golsefid

SMM

Bastani

. Dual centers Fuzzy Type-2 Clustering. 2013 Joint IFSA World Congress and NAFIPS Annual Meeting (IFSA/NAFIPS) [Internet]. 2013; 1215-20. Available from: https://api.semanticscholar.org/CorpusID:11042369.

Hassanien

Emary

Zawbaa

. Retinal blood vessel localization approach based on bee colony swarm optimization, fuzzy c-means and pattern search. J Vis Commun Image Represent [Internet].2015; 31: 186-96. Available from: https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S1047320315001182.

Golsefid

SMM

Fazel Zarandi

. Dual-centers type-2 fuzzy clustering framework and its verification and validation indices. Appl Soft Comput [Internet].2016; 47: 600-13. Available from: https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S156849461500318X.

Emary

Zawbaa

Hassanien

Schaefer

Azar

. Retinal vessel segmentation based on possibilistic fuzzy c-means clustering optimised with cuckoo search. In: 2014 International Joint Conference on Neural Networks (IJCNN). 2014. pp. 1792-6.

Roychowdhury

Koozekanani

Parhi

. Blood vessel segmentation of fundus images by major vessel extraction and subimage classification. IEEE J Biomed Health Inform.2015; 19(3): 1118-28.

Franklin

Rajan

. Computerized screening of diabetic retinopathy employing blood vessel segmentation in retinal images. Biocybern Biomed Eng [Internet].2014; 34(2): 117-24. Available from: https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S0208521614000205.

Melin

Sánchez

. Optimal design of type-2 fuzzy systems for  diabetes classification based on genetic  algorithms. Int J Hybrid Intell Syst.2021; 17: 15-32.

Rubio

Castillo

Valdez

Melin

Gonzalez

Martinez

. An Extension of the Fuzzy Possibilistic Clustering Algorithm Using Type-2 Fuzzy Logic Techniques. Xiong N, editor. Advances in Fuzzy Systems [Internet].2017; 2017: 7094046. Available from: doi: 10.1155/2017/7094046.

Moattar Husseini

Fazel Zarandi

Ahmadi

. Adaptive Type2-Possibilistic C-Means Clustering and Its Application to Microarray Datasets. Artif Intell Rev [Internet].2023 Mar; 56(10): 11017-52. Available from: doi: 10.1007/s10462-022-10380-5.

10.

Cordero-Martínez

Sánchez

Melin

. Hierarchical genetic optimization of convolutional neural models for diabetic retinopathy classification. Int J Hybrid Intell Syst [Internet].2022 Jan; 18(1–2): 97-109. Available from: doi: 10.3233/HIS-220004.

11.

Kaur

, Lalit &, Singh

. A method of disease detection and segmentation of retinal blood vessels using fuzzy c-means and neutrosophic approach. Imperial Journal of Interdisciplinary Research (IJIR).2016; 2(6).

12.

Duong

Nguyen

Ngo

Tinh

. On approach to vision based fire detection based on type-2 fuzzy clustering. In: 2011 International Conference of Soft Computing and Pattern Recognition (SoCPaR). 2011. pp. 51-6.

13.

Ashir

Ata

Salman

. FPGA-based image processing system for Quality Control and Palletization applications. In: 2014 IEEE International Conference on Autonomous Robot Systems and Competitions, ICARSC 2014. 2014.

14.

Gonzalez

Woods

. Digital image processing [Internet]. Upper Saddle River, N.J.: Prentice Hall. 2008. Available from: http://www.amazon.com/Digital-Image-Processing-3rd-Edition/dp/013168728X.

15.

Canny

. A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell.1986; PAMI-8(6): 679-98.

16.

Ashir

Eleyan

. A multi-resolution approach for edge detection using Ant Colony Optimization. In: 2015 23rd Signal Processing and Communications Applications Conference, SIU 2015 – Proceedings. 2015.

17.

Ashir

Piqueira

JRC

. Multilevel Thresholding for Image Segmentation Using Mean Gradient. Journal of Electrical and Computer Engineering. 2022; 2022.

18.

Hao

Changshun

Lei

. An improved method of image edge detection based on wavelet transform. In: 2011 IEEE International Conference on Computer Science and Automation Engineering. 2011. pp. 678-81.

19.

Kauppi

Kalesnykiene

Kämäräinen

Lensu

Sorri

Raninen

, et al. The DIARETDB1 Diabetic Retinopathy Database and Evaluation Protocol. In: British Machine Vision Conference [Internet]. 2007. Available from: https://api.semanticscholar.org/CorpusID:15483141.

20.

Muniyappan

Allirani

Saraswathi

. A novel approach for image enhancement by using contrast limited adaptive histogram equalization method. In: 2013 Fourth International Conference on Computing, Communications and Networking Technologies (ICCCNT). 2013. pp. 1-6.

21.

Dice

. Measures of the amount of ecologic association between species. Ecology [Internet].1945; 26(3): 297-302. Available from: http://www.jstor.org/stable/1932409.

22.

Jaccard

. The distribution of the flora in the alpine zone. New Phytol [Internet].1912; 11(2): 37-50. Available from: http://www.jstor.org/stable/2427226.

23.

Wang

Bovik

Sheikh

Simoncelli

. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing.2004; 13(4): 600-12.

24.

Ashir

Ibrahim

Abdulghani

Ibrahim

Anwar

. Diabetic Retinopathy Detection Using Local Extrema Quantized Haralick Features with Long Short-Term Memory Network. Int J Biomed Imaging. 2021; 2021.

25.

Faisal

Wahono

Purnama

IKE

Hariadi

Purnomo

. Classification of diabetic retinopathy patients using support vector machines (SVM) based on digital retinal image. J Theor Appl Inf Technol.2014 Jan; 59(1): 197-204.

26.

Lakshman

. Retinal image analysis using morphological process and clustering technique. Signal Image Process [Internet].2013; 4: 55-69. Available from: https://api.semanticscholar.org/CorpusID:14729428.

27.

Jahiruzzaman

Aowlad Hossain

ABM

. Detection and classification of diabetic retinopathy using K-means clustering and fuzzy logic. In: 2015 18th International Conference on Computer and Information Technology (ICCIT). 2015. pp. 534-8.

28.

Sopharak

Uyyanonvara

Barman

. Automatic exudate detection from non-dilated diabetic retinopathy retinal images using fuzzy c-means clustering. Sensors [Internet].2009; 9(3): 2148-61. Available from: https://www-mdpi-com-s.web.bisu.edu.cn/1424-8220/9/3/2148.

29.

Acharya

Lim

EYK

Chee

Tamura

. Computer-based detection of diabetes retinopathy stages using digital fundus images. Proc Inst Mech Eng H [Internet].2009; 223(5): 545-53. Available from: doi: 10.1243/09544119JEIM486.

30.

Vujosevic

Benetti

Massignan

Pilotto

Varano

Cavarzeran

, et al. Screening for diabetic retinopathy: 1 and 3 nonmydriatic 45-degree digital fundus photographs vs 7 standard early treatment diabetic retinopathy study fields. Am J Ophthalmol [Internet].2009; 148(1): 111-8. Available from: https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S0002939409001536.

31.

Jeyalaksshmi

Padmapriya

and MD and AA. Detection of Hard Exudate from Diabetic Retinopathy Image Using Fuzzy Logic. In: Peng Sheng-Lung and Son LH and SG and BD, editor. Intelligent Computing and Innovation on Data Science. Singapore: Springer Singapore. 2020. pp. 543-50.

32.

Prentašić

Lončarić

. Detection of exudates in fundus photographs using deep neural networks and anatomical landmark detection fusion. Comput Methods Programs Biomed [Internet].2016; 137: 281-92. Available from: https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S0169260716308033.

33.

Walter

Klein

Massin

Erginay

. A contribution of image processing to the diagnosis of diabetic retinopathy-detection of exudates in color fundus images of the human retina. IEEE Trans Med Imaging.2002; 21(10): 1236-43.

34.

Imani

Pourreza

. A novel method for retinal exudate segmentation using signal separation algorithm. Comput Methods Programs Biomed [Internet].2016; 133: 195-205. Available from: https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S0169260715301334.

35.

Cordero-Martínez Rodrigo, Sánchez

, and MP. Comparison of Image Pre-processing for Classifying Diabetic Retinopathy Using Convolutional Neural Networks. In: Abraham Ajith and Siarry P and PV and GN and CG and CO and HP, editor. Hybrid Intelligent Systems. Cham: Springer International Publishing. 2022. pp. 194-204.

36.

Abiyev

Altiparmak

. Type-2 Fuzzy Neural System for Diagnosis of Diabetes. Abdullah L, editor. Math Probl Eng [Internet]. 2021; 2021: 5854966. Available from: doi: 10.1155/2021/5854966.

Interval type-2 fuzzy approach for retinopathy detection in fundus images

Abstract

Keywords

1. Introduction

2. Theoretical background and motivation

2.1 Fuzzy c-means (FCM)

2.5 Canny edge detection

3.1 Features extraction

3.2 Blood vessel segmentation

4.1 Evaluation

Table 3 Comparisons with different methods

References

Table 3
Comparisons with different methods