Proliferative Diabetic Retinopathy Diagnosis: Comparison
Please note this is a comparison between Version 1 by Noor Ul Huda and Version 2 by Catherine Yang.

Diabetic retinopathy is one of the abnormalities of the retina in which a diabetic patient suffers from severe vision loss due to an affected retina. Proliferative diabetic retinopathy (PDR) is the final and most critical stage of diabetic retinopathy. Abnormal and fragile blood vessels start to grow on the surface of the retina at this stage. It causes retinal detachment, which may lead to complete blindness in severe cases. 

  • diabetic retinopathy
  • proliferative diabetic retinopathy
  • autonomous disease detection

1. Introduction

Diabetic retinopathy (DR) is a vascular complication of eye. It is the most severe abnormality among all diabetic eye diseases [1]. In DR, at first a lesion starts to appear on the retina, and then it results in the bleeding of blood vessels and capillaries on the surface of the retina. Due to the leakage of blood vessels, the oxygen supply to the retina decreases and as a consequence the brain stimulates the formation of new blood vessels to fulfill the requirement of oxygen. These new vessels may bleed and cause the detachment of the retina and ultimately loss of vision [2][3][2,3]. Out of all DR patients, few suffer from proliferative diabetic retinopathy (PDR), but if it is not diagnosed in a timely manner, the disease may cause severe destruction. Neovascularization appears as a tortuous collection of blood vessels and is quite destructive because these vessels grow abnormally out of the retina into the clear vitreous gel [4][5][4,5]. Therefore, vessels grow beyond the supporting structure of the retina and they are very disposed to bleeding, particularly when they arise near the optic disc. Even a small rise in blood pressure can cause hemorrhages in this case. If bleeding appears in the vitreous humor it can affect the visual system. If this bleeding becomes extensive, it results in a painless and rapid blackening of the vision. Neovascularization [6] is divided into two types [7]:
  • Neovascularization on disc (NVD): If the new vessel formation occurs within one disc diameter of the optical disc then this is categorized as NVD or neovascularization on disc
  • Neovascularization elsewhere (NVE): If new vessel formation occurs elsewhere on the surface of the retina, then this is called neovascularization elsewhere (NVE).

2. Machine-Learning-Based Algorithms

Gotman et al. [8][11] proposed a method for the detection of new vessels on the disc using a support vector machine (SVM) classifier. They used a watershed transform combined with 2D Gaussian for the segmentation of blood vessels. Fifteen features were used for the classification of the SVM classifier and they achieved an area under receiver operating characteristics curve (ROC) of 0.909. They came up with a feature set to detect neovascularization but they limited their scope to NVD only. An amplitude modulation–frequency modulation (AM-FM)-based method was proposed by Agurto et al. [9][12]. They followed a top-down approach for the detection of NVD. They came up with a sound set of features and used K-means clustering on these feature sets. Jelinek et al. [10][13] proposed a new technique for the detection of proliferative diabetic retinopathy from angiograms. They used 27 labeled images and achieved an accuracy of 0.90 with the selection of six features. Derivatives of the Gaussian wavelet were used in their work for the segmentation of blood vessels. Mudigonda et al. [11][14] proposed a method for the detection of neovascularization in retinal fundus images using fractal analysis. The proposed technique used colored fundus images as the input, followed by ROI extraction, i.e., extraction of the region around the optic disc. The green channel was extracted from the ROI region to obtain the maximum information from the vessels. The vessels were extracted using a Gabor filter, and the resulting magnitude image was converted to a binary image. The resultant image was analyzed using a fractal analysis box-counting method that identified vessel bundles near the optic disc region, i.e., neovascularization in the optic disc region (NVD). Among ten images, five with neovascularization had a fractal mean value of 1.66 and five images with no neovascularization in the optic disc region resulted in a fractal mean value of 1.58. Saranya et al. [12][15] used the fuzzy C-means (FCM) technique for blood vessel segmentation. They used a set of features that included the gradient, gradient variation, gray-level coefficient of variation, moment invariant-based features, and tortuosity for a k-nearest neighbor (KNN) classifier. They achieved an accuracy of 96.5% on the DRIVE and MESSIDOR datasets. A research methodology for the automated detection of neovascularization for PDR proposed by Sohini Roy Chowdhury et al. [13][16] describes a technique to detect neovascularization from fundus images and classify it as neovascularization in the optic disc region (NVD) or neovascularization elsewhere (NVE). The green plane is extracted from the input image and normalized in the [0,1] intensity range. The region of interest (ROI) is extracted for both types of neovascularization, leading to vessel detection from both ROIs. Textural, structural, and intensity-based features are used to classify NVD and NVE. The proposed method was trained and tested on 40 images (30 normal, ten with PDR) from the STARE database and 17 images from a local dataset. Accuracies of 87.6% and 92.1% were obtained for NVD and NVE, respectively. Lee et al. [14][17] proposed a new vessel-detection method that includes statistical texture analysis (STA), high-order spectrum analysis (HOS), and fractional analysis. They used a total of 137 images in their work and achieved an area under the curve of 99.3%. A method based on the following-the-line approach for the segmentation of vessels was proposed by R.A. Welikala et al. [15][18]. They used two-line approaches and two different sets of features for the retraction of true abnormal blood vessels. They used 60 images from the MESSIDOR dataset for evaluation purposes and classified these images on the basis of an SVM classifier. They achieved an area under the curve (AUC) value of 0.96. Shuang Yu et al. [16][19] proposed a novel technique for the automation of neovascularization in the optic disc region (NVD). A fundus image is pre-processed followed by the application of a Gabor filter to extract blood vessels. Twenty-one texture-based and vessel-based features are extracted to classify an image as normal or NVD using support vector machines (SVM). Sixty-six retinal images (15 NVD, 50 normal) were extracted from the globally available MESSIDOR, HRF, and DIARETDB0 datasets to test and train the proposed technique. A sensitivity of 15/16 and specificity of 47/50 were achieved using the proposed methodology. In 2016 research was conducted by Diego F. G. Coelho [17][20] that aimed to detect NVD from fundus images. In the proposed methodology, fundus images are analyzed by calculating the gradient magnitude of the Fourier power spectrum followed by extraction of the angular spread. Entropy and spatial variance are used to categorize an image as a normal image or image with NVD using a linear statistical classifier. An accuracy of 100% was achieved when the proposed technique was tested on ten images (five normal, five NVD) extracted from the MESSIDOR database. Akram et al. [7] presented a method for the detection of PDR. Their proposed method extracted a number of features based on vascular patterns for the proper representation of normal and abnormal vessels. A modified m-mediods-based classifier was used for the proper discrimination of abnormal vessels from normal ones. Another machine-learning-based technique for automated NVD detection was proposed by Shuang Yu et al. [18][21]. The proposed algorithm takes a fundus image and extracts the ROI, i.e., the disc region. Vessels are extracted using multilevel Gabor filters. A feature vector with 42 features (morphological and texture based) is extracted from both normal and NVD images, followed by a reduction in size to eighteen. A reduced feature vector was used to train and test 424 (134 NVD, 290 non-NVD) retinal fundus images. An accuracy of 95.23% was observed. Christodoulidis et al. [19][22] proposed a novel technique for the detection of NVD from fundus images. The proposed research states that NVD detection from retinal fundus images can be improved by adding a second-order statistical feature to the existing feature set containing structural, vessel-based, and intensity-based features. The image is pre-processed, followed by vessel detection. The vessel junctions are extracted by applying the Tensor voting technique, which highlights the local maxima, indicating the junctions of vessels. The suggested feature addition to the feature vector improved the sensitivity to 0.84. Mona Leeza et al. [20][23] proposed an algorithm in 2019 to detect the severity level of diabetic retinopathy using the bag-of-features approach. The algorithm is composed of five phases starting from local feature extraction from retinal images using SURF. K-means clustering is used to cluster the extracted features for dictionary generation. The algorithm proceeds by max pooling to accumulate features followed by the construction of histograms of oriented gradient (HoG). SVM and artificial neural networks are used to classify the retinal image as normal, mild NPDR, moderate NPDR, severe NPDR, and PDR. The algorithm resulted in 95.92% and 98.90% sensitivity and specificity, respectively. Research conducted by Lei Zhang et al. [21][24] described another algorithm to screen for PDR using a modified matched-filter approach. In the proposed technique, the result from Gaussian is proceeded by subtraction of the mean to eliminate the false positives that occur due to step edge noise. The accuracy of the algorithm was evaluated on the ZUEYE database, which resulted in an accuracy of 95%.

3. Deep-Learning-Based Algorithms

A technique was proposed in 2019 [22][25] to detect and categorize diabetic retinopathy using a deep convolution neural network (CNN) of five layers. The algorithm starts with pre-processing of the images, followed by an ensemble CNN model. The ensemble CNN model is composed of five deep CNN models, i.e., Resnet50, Inceptionv3, Xception, Dense121, and Dense169. The CNN model classifies the input image as normal, mild, moderate, severe, or PDR. The algorithm was tested on the Kaggle dataset, composed of 35126 colored retinal images. Sixty-four percent of images in the dataset were used for training, 20% images were used for testing, and 16% were used for validation. Specificities of 0.40, 0.99, 0.95, 0.98, and 0.99 were observed for each category, respectively. In 2022 [23][26], a neural-network-based framework was proposed using optical coherence tomography (OCT) images. The proposed algorithm classified an OCT image as normal or diseased using 3D feature extraction. Initially, segmentation was performed to extract 12 layers from the input image followed by feature extraction, i.e., thickness and angle calculation. The extracted features were then fused and passed to the neural network to make a decision, yielding an accuracy of 96.61%. Ayesha et al. [24][27] proposed three deep neural frameworks for diabetic retinopathy grading using retinal fundus images. The first framework used cascaded architecture to grade a retinal image among five grades of PDR using a three-layer CNN architecture. The second framework utilized the hue saturation value (HSV), red green blue (RGB), and normalized input image to apply ensemble-based architecture, where the final results were deduced using average pooling from each CNN model. The third framework incorporated a long short-term memory (LSTM) module to enhance the network memorizing capabilities. The EyePACS dataset containing 88,702 retinal images was used to train and test the proposed framework. Among all, the ensemble-based architecture outperformed, resulting in an accuracy of 83.78%. Another framework proposed by Tang et al. [25][28] segmented and localized neovascularization using a deep learning architecture. The proposed algorithm starts with image pre-processing followed by dividing the input image into non-overlapping patches. To train the neural network, the ground truth containing neovascularization in each patch was passed as a training dataset, which classified each pixel of the patch as neo or non-neo. The dataset was divided into validation, training, and testing sets, yielding an accuracy of 0.9948 on a dataset with 50 images. This research was further extended [26][29] using transfer learning on pre-trained models that included AlexNet, GoogLeNet, ResNet18, and ResNet50 pre-trained on ImageNet. Ground truth patches were used for training these models, followed by testing the models. Another module utilized the pre-trained model for feature extraction followed by classification using SVM. In addition to using pre-trained models separately, a combination of ResNet and GoogleNet was proposed that yielded the highest accuracy of 0.9157. Another algorithm [27][30] utilized ResNet to detect retinal neovascularization from retinal fundus images. The proposed framework pre-processed the input image to enhance the contrast and remove noise followed by training ResNet. The proposed model was trained on 3662 retinal images from a local dataset containing healthy images (Label 0), neovascularized images (Label 2), and diabetic retinopathy images (Label 1). Due to the residual properties of ResNet, the model resulted in an accuracy of 0.88 on 1992 retinal fundus images.