Analysis of retinal images is used for diagnosing and treatment of several diseases related to the eye such as DR and glaucoma. The benefits of using computerized retina image analysis are many such as (i) manual diagnosis is usually performed by analyzing the images from a patient, as not all images show signs of DR; (ii) there are insufficient number of healthcare specialists to perform eye examination for huge population of diabetics, (iii) computer-assisted diagnosis of retinal images; used in large screening of diabetics for retinopathy is economical.

Every part of our human body thrives on nourishment derived from the circulating blood. Whenever blood supply to a particular tissue is compromised, the vitality of that tissue is threatened. Diabetes mellitus damages the small blood vessels of the body. As the retina is nourished by small microscopic blood vessels, DR is the damage to the retina due to the diseased blood vessels. Damage to the blood vessels cause the major type of DR [1]. Hemorrhages appear either as small red dots or blots indistinguishable from micro-aneurysms or as larger flame-shaped as shown in Fig. 1.

Early detection of lesions is very important for ophthalmologists in clinical diagnostic and successful treatment. So this paper aims at building a system for early detection and classification of major type of DR.

Hemorrhages in retinal images can be difficult to identify because regions that appear to be small red dots could really be normal. In addition, the blood vessels and fovea are difficult to distinguish from red dots in the retinal image. These can lead to ophthalmologists being misled. The proposed method in this paper is to detect normal objects such as blood vessels, optic disc and fovea and which aids in the detection of red dots. These objects are masked and subsequently hemorrhages are identified. This type of DR are detected and classified using three steps: (i) segmentation of the dark regions in the image using adaptive thresholding algorithm, followed by top hat morphological technique (ii) statistical features and pattern recognition techniques are extracted from each region (iii) SVM classifier is used to classify the regions either as healthy or non-healthy.

The main contributions of this paper are summarized as follows:1)

A new technique to segment normal objects to distinguish them from hemorrhages.

This paper is organized in five sections. In Section 2 the related work is discussed. The technique for segmentation of hemorrhages is described in Section 3. In Section 4 feature extraction is presented. Experimental results and comparative analysis with some existing methods are presented in Section 5.In Section 6 the conclusion is presented.

Different techniques have been used for the detection of hemorrhages in retinal images. Mamilla et al. [2] presented a new technique to extract small red dots based on phase congruency and mathematical morphology. Adem et al. [3] proposed a technique to extract dark regions by iterative thresholding approach based on firefly and particle swarm optimization. Thresholding, morphology and geometrical features are used to detect lesions from retinal images [4].

Bae et al. [5] employed a method for the extraction of candidate dark lesions using a circular template matching with cross correlation. Then the candidates are segmented by two methods of region growing: local threshold and adaptive seed region growing segmentation.

Zhou et al. [6] used a method based on super-pixel to extract the candidate hemorrhages. This is followed by extracting features from these candidates. Subsequently, Fisher Discriminant Analysis classifier is applied to classify the candidates as either healthy or red dots.

Chowdhury et al. [7] applied a new technique to segment the hemorrhages using K means clustering in fundus images and is followed by Random Forest Classifier to classify the retinal images. Hatanaka et al. [8] suggested a rule-based method and three Mahalonobis distance classifiers to classify the candidate hemorrhages. Tang et al. [9] utilized a splat feature classifier for lesions detection in fundus images. A supervised method was used for partitioning image into non-overlapping segments called clusters, consisting of similar color pixels. Features are extracted from these clusters which describes its characteristics. Optimal feature subset is selected by the wrapper approach.

Orlando et al. [10] in this research employed a new technique to extract dark lesions in retinal images based on combining both deep learned and domain knowledge. Features learned by a convolutional neural network (CNN) are augmented by incorporating hand crafted features.

Preprocessing of retinal images is initially performed by Aziz et al. [11]. After that, red dots are segmented using a smart windowing-based adaptive threshold. Finally, the regions conventional and handcrafted features are extracted and classified using a SVM.

A new convolutional neural network (CNN) architecture is presented to detect DR. Features are extracted from the identified blots indistinguishable from micro-aneurysms using a modified pre-trained CNN model. Finally, the multi-logistic regression controlled entropy variance methodology is used to select the best features [12]. For extracting candidates for large red-lesions, Asiri et al. [13] developed a new deep learning-based technique. They have used two different candidate generating algorithms, one for small red lesions and the other for large red lesions, as well as two different stream structures based on candidate type. Also, they utilized the VGG-16 model with ROI pooling, regression, and classification layers for each stream. Khojasteh et al. [14] used probabilistic output from CNN to automatically and simultaneously detect lesions in retinal images. Image pre-processing, training data refining, and object detection employing a convolutional neural network with label smoothing are the three parts of the suggested system to extract hemorrhages used by Huang et al. [15].

The methods described in the literature are focused on identifying a certain type of dark lesion. However, the blood vessels and fovea inside the retinal image have same the features with hemorrhages that lead to false detection. So the proposed strategy for locating hemorrhages is based on the following steps: Locating dark areas in the remainder of retinal image after masking segmented regular objects (vessels, optic disc and fovea) from retinal image.

Selecting the clinical descriptive features to describe these candidate areas.

The segmentation of dark regions in fundus retinal images is a difficult step because the normal objects inside retinal images such as blood vessels and fovea have the same color, shape and size features as the hemorrhages. Therefore, as a preprocessing step, the blood vessels, optic disc and fovea are segmented and masked. The sequence of steps is as shown in Fig. 2.

The purpose of this process is to detect the candidate hemorrhages from the retinal image. The first step is the enhancement of the image using Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve the contrast between hemorrhages and background. A local adaptive thresholding technique is applied first to separate hemorrhages image objects from the background based on the local mean intensity. (Subsequently, black morphological top hat is applied to the image in two steps (i) closing is done on the image to obtain the candidates dark regions in the image using disk shape structuring element (ii) subtracting the input image from the output of the previous step.(4) Closing(A,SE)=(A⊕SE)⊖SE (5) Tb=Closing(A,SE)−A

Here A is the input image in green channel and SE is symmetrical disc structuring element with radius 12. Fig. 5b shows the binary image of the hemorrhage candidates in white color and the background in black color.

Not all the extracted dark regions are hemorrhages. The extracted dark areas need to be checked carefully in order to decide; if they are hemorrhages or normal background. This task is achieved by selecting a set of features for each region. The technique for selecting features is based on grouping features into two main categories: One is based on the ophthalmologist's recognition of lesions using color and size; and the other is based on pattern recognition techniques by extracting texture, dominant color and wavelet features.

SVM are supervised learning models with associated learning algorithms that analyze data and recognize patterns in the input data. In our approach three different types of SVM are applied; Linear SVM, Quadratic SVM and Cubic SVM that takes a set of input data and predicts the output, which forms two possible classes.

In our work, we have extracted 3386 candidate dark regions from the dataset. Each region is represented with 23 extracted features. The experiment is conducted by varying the number of candidate areas. The candidate areas (testing data set size) are 30%, 40%, 50% and 60%.

The proposed technique has been experimented on the publicly available database DIARETDB1 [20].This dataset consists of 89 fundus images with 50° field-of-view. Each image has a resolution of 1500 × 1152 pixels. Further, the ground truth label for segmentation of optic disc and hemorrhages are available from the dataset. The dataset contains 59 abnormal images with the labels.

To measure the performance of the proposed method, experiments are done using proposed and state-of-the-art methods to segment optic disc and classify the hemorrhages. Metrics such as the average sensitivity, specificity, average predictive value and F-score are taken to quantify the performance.(16) Sensitivity(Se)=TP(TP+FN) (17) Specificity(Sp)=TN(TN+FP) (18) Accuracy(Acc)=(TP+TN)(TP+FP+TN+FN) (19) PPV=TP(TP+FP) (20) F−score=2TP(2TP+FP+FN)

The results of candidate window selection and extraction of the optic disc on two images using variance are shown in Fig. 6. Tab. 2 shows the performance of the proposed method in three terms and is compared with state-of-art-methods for segmentation optic disc.

The overall algorithm for segmentation of optic disc is given below

Extensive experimentation with different thresholds on various parameters are used, and found the best threshold value to extract optic disc is 160.

From Tab. 2 it is evident that the proposed method achieves better performance than the other methods with respect to average sensitivity for segmentation of optic disc. Fig. 7 shows green channel retinal images after masking normal objects and the candidate region of hemorrhages.

The overall algorithm for detection and classification of hemorrhages is given below

The performance of the proposed method for detection hemorrhages are tabulated (Tabs. 3 to 5) for linear SVM, quadratic SVM and cubic SVM respectively and compared with the other state-of-art-methods as shown in (Tab. 6). Figures 8 and 9 show the graph ofcomparison on different techniques for detection of haemorrhages.

From Tabs. 3 to 5 we observe that Linear SVM classifier gives better results than Cubic SVM and Quadratic SVM with respect to sensitivity and accuracy and with respect to specificity Quadratic SVM gives better results as compared to other SVMs.

An efficient framework for the early diagnosis of DR diseases such as hemorrhages has been proposed in this paper.We have developed a new technique to detect and classify hemorrhages in retinal images. The proposed method consists of four steps; (i) normal objects in retinal images are detected and masked because these objects have the same features with hemorrhages that lead to misdiagnosis. (ii) The dark regions are extracted by using adaptive thresholding and top-hat morphological techniques. (iii) Statistical and pattern recognition features are extracted from each region (iv) SVM is used to classify the candidate dark regions as either hemorrhages or healthy.

We have achieved 90.4%, 89.5%, 90.1% and 92.7% for sensitivity, specificity, accuracy and F-score respectively. Further, the proposed method is compared with the state-of-art-methods.

In the future, we intend to collect a private dataset that contains more than a thousand images to apply deep learning classification techniques. Also we plan to detect and classify other types of DR diseases such as Micro-aneurysms.