The substantial vision loss due to Diabetic Retinopathy (DR) mainly damages the blood vessels of the retina. These feature changes in the blood vessels fail to exist any manifestation in the eye at its initial stage, if this problem doesn’t exhibit initially, that leads to permanent blindness. So, this type of disorder can be only screened and identified through the processing of fundus images. The different stages in DR are Micro aneurysms (Ma), Hemorrhages (HE), and Exudates, and the stages in lesion show the chance of DR. For the advancement of early detection of DR in the eye we have developed the CNN-based identification approach on the fundus blood lesion image. The CNN-based automated detection of DR proposes the novel Graph cutter-built background and foreground superpixel segmentation technique and the foremost classification of fundus images feature was done through hybrid classifiers as K-Nearest Neighbor (KNN) classifier, Support Vector Machine (SVM) classifier, and Cascaded Rotation Forest (CRF) classifier. Over this classifier, the feature cross-validation made the classification more accurate and the comparison is made with the previous works of parameters such as specificity, sensitivity, and accuracy shows that the hybrid classifier attains excellent performance and achieves an overall accuracy of 98%. Among these Cascaded Rotation Forest (CRF) classifier has more accuracy than others.

Retinopathy is a chronic visual impairment [

Micro aneurysms (Ma) [

Through fundus imaging [

The classification is done by our proposed hybrid classification method as K-NN, SVM with PCA, and Cascaded Rotation Forest. These methods were separately used for classification but in this developed approach all three classifiers were hybridized and proceeded for classification and yielding high accuracy of 98%. Thus, we can detect all DR-related visual impairments more accurately and with no delay time.

This research paper is then followed by: (related studies) literature on the existing works about the detection of diabetic retinopathy and its drawbacks, then the followed segment 3 made suggested research on the detection of diabetic retinopathy, and part 4 narrates the experimental analysis and outcomes of the suggested detection method and last but not least the last segment holds the conclusion part of this paper.

In 2017, Ghosh, R et al. proposed the automatic detection and classification technique for different stages of Diabetic Retinopathy using CNN networking, the input color fundus retinal image is denoised by CNN’s six networking layers with this Micro-aneurysm hemorrhage stages of DR are classified and identified [

In 2020, Shaban, M et al., developed the specified CNN for the screening and staging of DR [

In 2017, Soomro, et al. proposed an extraction method for the analysis of DR using Basic Filtering Schemes from the retinal blood vessel [

In 2021, Gayathri, S et al. proposed a research paper on the classification of Diabetic Retinopathy using both Multipath CNN and the machine learning classifiers [

In 2022 Jaichandran, R et al. proposed a paper to detect diabetic retinopathy using a CNN in computer vision-based technique [

In 2021, Yadav, et al. developed the Color locus detection approach that employs networking such as ANN, CNN, and AI to identify early microaneurysms in DR patients [

The interior part of an eye is imaged through a fundus camera and this image is taken as an input image. The pre-processing is done from the acquired input image [

The functional pattern of the developed Graph theory-based segmentation is depicted in

Initially, preprocessing [

Steps of Wiener filtering:

Taking Fourier transformations for the blurred low quality input image.

Determination of output function from the product of Wiener filtering and Fourier transform.

D (m,n)–Applying Fourier transform to a blurred image.

I^{T}

After Weiner Filtering, Inverse Fourier transform is taken to transform to produce a deblurred image Y(i, j).

The image processing stages are as follows, 1) Image sizing, 2) Green channel extraction, 3) Image contrast enhancement by Histogram Equalization, and 4) Extraction of the blood vessel is the most significant for pre-processing the image.

The input fundus image is resized. Red, green, and blue are the three primary channels for extraction [

For each distinct part of the image, several histograms are computed and these are used to modify the contrast adjusting the pixel brightness value. The sectioned image is enhanced depending on the boundary of each section and the contrast-enhanced image which is shown in

The same featured elements are grouped, and this similar Structuring element (S) is used for different processing such as dilating, eroding, and closing the image (I_{C}) and the above processing-related function called morphological closing.

In

Graph cut segmentation is a recurrent technique for extracting the specific areas of an image that represent diabetic retinopathy lesions and the segmented image using the graph cut technique is shown in

In Graph cut segmentation-based approaches, each pixel is treated as a node in the segmented graph, and superpixels are formed and specified across the graph by the cost function reduction. In the graph, we removed the minimum sum of edge weights to divide it into two groups, U and V. Similarly, we determine the minimum cut, which is a maximum flow.

Steps:

Differentiating the image’s cluster points.

Evaluating the length of distinct Unique feature.

Identifying the index for each distinct feature.

Clusters are divided into groups depending upon their distance from each other.

The image is clustered depending on the maximum flow cut.

Graph segmentation produces Volume (U) and Volume (V), which represent the sums of costs of edges connected to U and V, respectively.

From feature extraction, some of the features like statistical, texture, HOG & SIFT are extracted, [

Mean: In image processing, the pixel value in the image is the prime factor, in which the overall add of pixel value or brightness value (B) in a segmented image per total count of pixel (C) is called as mean (μ) of the sliced image.

Variance: The measurement of differences from the pixel’s mean value is called the variance of the segmented image, this is mentioned as (X). Its square root (X^{2}) is defined by standard deviation

Steps:

Calculate the average value of the specific segmented image.

Finding the difference of segmented image mean value from every pixel in an image.

Calculate the squaring of every deviation away from the mean of the segmented image.

Find the total addition of each one squaring value.

Divide the entire number of pixels by the addition of squares.

Probability Density Functions (PDF) are used to extract the texture characteristics. The Probability density is computed using GLDM from different images [

The following are the steps involved in GLDM feature extraction:

Accessing the image that has been segmented [

Calculating image dissimilarity.

I–is the input images

Δi, Δj–spotted position of image while incrementation.

The probability density function of d (m, n) is used to extract the texture characteristics in four different directions such as (0°, 45°, 90° and 135°).

Creating a feature vector by joining the derived characteristics such as difference, average, and variance.

The characteristics of the lesion play a key role in proper disease prediction of DR, and this may be handled via HOG feature extraction. The HOG feature descriptor involves the following steps: 1) Image is separated into blocks 2) Using the mask for both longitudinal and transverse directions and calculating the gradation level 3) Measuring the number of times gradient direction appears. 4) The final values are standardized. 5) PCA is taken to reduce the number of features.

SIFT is a seven-stage procedure that results in a huge number of properties such as illumination constant, scaling constant, and rotation constant in the brightness of the segmented image. As a result, this might work for a wide range of image resolutions [

After the feature extraction, the fundus images are processed into a classification using different Machine Learning classifiers such as the K-NN classifier, SVM with PCA, and CRF hybrid classifiers. The classifiers are trained and validated using the extracted features. About 500 fundus image data sets are taken for training and 600 fundus data sets for testing the lesions.

Once the images are acquired and stored for the training process and then the images are classified into various categories which must be identical to the acquired image. Calculating the image pixel’s neighborhood relationship and similarity from each K-NN, Euclidean, and Manhattan distance. This finding plays a typical role, but it’s a large scale for finding the nearest neighbor distances.

We adopted a rotating forest classifier because of the benefits of integrating the overall prediction of numerous base classifiers, lowering variance and error in the end. The following are the steps involved in creating a CRF classifier:

N: It is the numerical value of how many classifiers are used in classification (C_{1}, C_{2},… ,C_{N})

T: It is the training fundus data set.

L: It represents the label of class.

d: The classifiers provide a probability per each data set.

for i = 1: N

Assigning base classifiers as decision trees: C_{i}

Build training data for the decision tree-based classifiers

- The separate feature set S into M subsets: S_{ij}

for j = 1:M

for the features in S_{i,j}

Discard random subsets and select bootstrap sample from T_{i,j} to get T’_{i,j}

Apply PCA to T’_{i,j} and extract principle components

end for

construct the rotation matrix R_{i,}

Build classifiers C_{i,} using (TR_{i}, L)

end for

Based on the largest confidence level for an input x, assign it to a class,

Statistical learning is a theory that analyses the law of machine learning in the context of a minimal training sample. It introduces the Support Vector Machine, a novel classification approach with maximal hyper-plane edge separation. This section delves into the statistical analysis approach known as Principal Component Analysis (PCA). This strategy is utilized to master the principal contradiction of things. It may extract the main components from a variety of sources and show their essence.

The proposed research work is to detect the stages of Microaneurysms, Hemorrhages, and Exudates of lesions by which the DR is diagnosed. The methodology used here is to expand the blood vessels and conduct Graph cut segmentation to the background and foreground pixel placement. The performance of the segmentation is measured by the accuracy rate. It shows a solution to the binary problem in which the segmentation of graph cuts is used to detect exudates regions.

S.no | Classifier | Input | Training data set | Testing data set | Accuracy |
Sensitivity |
Specificity |
||
---|---|---|---|---|---|---|---|---|---|

No of the input data set | No. of correct |
No of inputs | Test output | ||||||

1. | K-NN | Normal | 250 | 250 | 200 | 185 | 91 | 85 | 91.5 |

DR | 310 | 310 | 400 | 340 | 85.5 | ||||

2. | SVM with PCA | Normal | 250 | 250 | 200 | 194 | 95.5 | 95 | 96.09 |

DR | 310 | 310 | 400 | 375 | 94.5 | ||||

3. | Cascaded Rotation Forest | Normal | 250 | 250 | 200 | 197 | 97.9 | 98 | 97.9 |

DR | 310 | 310 | 400 | 394 | 98 |

The graphical representation of both the normal and the disease-funded image is given by training the input images and then testing them using a trained data with a better accuracy rate. The normal fundus image is tested with 98% accuracy and the DR disease funded image is tested with an average of 98% accuracy.

Based on the following KNN classifier, Support Vector Machine (SVM) with PCA and Cascaded Rotation Forest classifiers, the performance parameters like sensitivity and specificity are graphically figured out in

Through the analysis of three classifiers, the spotting and categorization of Microaneurysms, Hemorrhages, and Exudates stages in DR were identified inevitably and its final resultant comparison is depicted in

From the KNN and SVM approach along with PCA classifier analysis, it identifies exudates condition is normal, but in CRF classifier output is positive for exudates, and this CRF classifier rectification is accurate while comparing with other classifiers which are demonstrated in Class I of

Sl.no | Methods | Lesions | Sensitivity | Accuracy (%) |
---|---|---|---|---|

1 | Gradient weighting technique and an iterative thresholding approach. | Blister aneurysms or microaneurysms | 0.815 | - |

2 | Patch-based fully convolutional neural network with batch normalization layers and Dice loss function. | Blister aneurysms or microaneurysms | 0.391 | |

3 | Toboggan segmentation algorithm. | blood loss or Hemorrhage | 0.911 | 93% |

4 | Circular Hough Transforms (CHT), contrast limited AHE (CLAHE), Gabor filter, and thresholding. | Dense exudates and Hemorrhages | - | - |

5 | Region growing and segmentation stage based on fuzzy c-means clustering. | Hemorrhage | 0.894 | - |

6 | Image Processing Technique. | Exudates and Hemorrhage | 0.962 | 92% |

7 | Minimum Intensity Maximum Solidity (MinIMaS) overlap algorithm | Microaneurysms, Hemorrhages, Exudates | 0.891 | 91% |

8 | Automated Fuzzy Inference System (FIS). | Exudates | 0.901 | 94% |

In this research paper, the implemented hybrid classifier of accurate cross-validation has the capability of detecting the DR at its initial level of abnormality through the identification of Microaneurysms, Hemorrhages, and Exudates in the fundus image lesion. In this, the hybrid classifiers were used to examine the intense level of DR from the performance parameters carried out through a comparison with different classifier methodologies. Based on the previous research works justification of all stages, severity, and detection of DR from the lesion is not possible for all the methodologies. The proposed algorithm of specific Graph theory segmentation and hybrid classification processing mainly focuses on detecting the DR of all lesion stages through the identification of Microaneurysms, Hemorrhages, and Exudates. Therefore, the hybrid K-NN, SVM, and Cascaded Rotation Forest classifiers enact comparison and achieve the accuracy of 98%. Among these Cascaded Rotation Forest classifier plays a vital role in terms of accuracy as it is a perfect automated DR classifier for a distinct stage of lesions. This work may have a future extension in the area of the proliferative stage of DR while detecting the DR stages of the lesion and the detection of DR in an early stage may be carried out by modern OCT scan methods based on hybrid deep learning methods.

We would like to thank the supervisors and the anonymous referees for their kind help in this research.