Automatic Diagnosis of Glaucoma from Retinal Images Using Deep Learning Approach PDF

diagnostics Article Automatic Diagnosis of Glaucoma from Retinal Images Using Deep Learning Approach Ayesha Shoukat 1 , Shahzad Akbar 1, * , Syed Ale Hassan 1 , Sajid Iqbal 2, * , Abid Mehmood 3 and Qazi Mudassar Ilyas 2 1 Department of Computer Science,...

diagnostics Article Automatic Diagnosis of Glaucoma from Retinal Images Using Deep Learning Approach Ayesha Shoukat 1 , Shahzad Akbar 1, * , Syed Ale Hassan 1 , Sajid Iqbal 2, * , Abid Mehmood 3 and Qazi Mudassar Ilyas 2 1 Department of Computer Science, Riphah International University, Faisalabad Campus, Faisalabad 44000, Pakistan 2 Department of Information Systems, College of Computer Sciences and Information Technology, King Faisal University, Al Ahsa 31982, Saudi Arabia; [email protected] 3 Department of Management Information Systems, College of Business Administration, King Faisal University, Al Ahsa 31982, Saudi Arabia * Correspondence: [email protected] (S.A.); [email protected] (S.I.) Abstract: Glaucoma is characterized by increased intraocular pressure and damage to the optic nerve, which may result in irreversible blindness. The drastic effects of this disease can be avoided if it is detected at an early stage. However, the condition is frequently detected at an advanced stage in the elderly population. Therefore, early-stage detection may save patients from irreversible vision loss. The manual assessment of glaucoma by ophthalmologists includes various skill-oriented, costly, and time-consuming methods. Several techniques are in experimental stages to detect early-stage glaucoma, but a definite diagnostic technique remains elusive. We present an automatic method based on deep learning that can detect early-stage glaucoma with very high accuracy. The detection technique involves the identification of patterns from the retinal images that are often overlooked by clinicians. The proposed approach uses the gray channels of fundus images and applies the data augmentation technique to create a large dataset of versatile fundus images to train the convolutional neural network model. Using the ResNet-50 architecture, the proposed approach achieved excellent results for detecting glaucoma on the G1020, RIM-ONE, ORIGA, and DRISHTI-GS datasets. We obtained a detection accuracy of 98.48%, a sensitivity of 99.30%, a specificity of 96.52%, an AUC of 97%, Citation: Shoukat, A.; Akbar, S.; and an F1-score of 98% by using the proposed model on the G1020 dataset. The proposed model may Hassan, S.A.; Iqbal, S.; Mehmood, A.; help clinicians to diagnose early-stage glaucoma with very high accuracy for timely interventions. Ilyas, Q.M. Automatic Diagnosis of Glaucoma from Retinal Images Using Keywords: glaucoma; fundus images; deep learning; early-stage detection; augmentation Deep Learning Approach. Diagnostics 2023, 13, 1738. https://doi.org/ 10.3390/diagnostics13101738 Academic Editor: Jae-Ho Han 1. Introduction The major components of the human eye involved in vision are the cornea, pupil, Received: 6 March 2023 iris, lens, retina, optic nerve, and tears. The iris is located between the cornea and the Revised: 4 May 2023 lens and controls the light. The retina receives the light and transfers it to the brain for Accepted: 6 May 2023 recognition by converting it into electrical signals. At the backside of the eye is a nerve Published: 14 May 2023 known as the optic nerve, which comprises 1 million nerve fibers of the retinal ganglion cells. The primary function of this nerve is to transfer visual signals from the retina to the occipital cortex. Copyright: © 2023 by the authors. The human eye contains a fluid known as aqueous humor, which is continuously Licensee MDPI, Basel, Switzerland. recycled. An obstruction in the drainage of aqueous humor leads to increased intraocular This article is an open access article pressure (IOP). Consequently, the retina and optic nerve are damaged, which may lead to distributed under the terms and vision loss. This is partly due to the degeneration of ganglion cells in the retina [2,4]. conditions of the Creative Commons The loss of optic nerve fibers changes the shape of the optic disc (OD) towards an increase Attribution (CC BY) license (https:// in the cup-to-disc ratio (CDR), which is an early sign of glaucoma. The anatomy of the creativecommons.org/licenses/by/ eye is depicted in Figure 1. The visual loss in glaucoma is due to damage to the retinal 4.0/). Diagnostics 2023, 13, 1738. https://doi.org/10.3390/diagnostics13101738 https://www.mdpi.com/journal/diagnostics  Diagnostics 2023, 13, x FOR PEER REVIEW 2 of 21 Diagnostics 2023, 13, 1738 in the cup-to-disc ratio (CDR), which is an early sign of glaucoma. The anatomy of2the of 17 eye is depicted in Figure 1. The visual loss in glaucoma is due to damage to the retinal ganglionic cells [7,8]. The alterations in the visual field scope are essential for diagnosing glaucoma ganglionic. Figure cells [7,8].2 The shows the enlarged alterations CDR in the in an visual eyescope field with are glaucoma. essential for diagnosing glaucoma. Figure 2 shows the enlarged CDR in an eye with glaucoma. Diagnostics 2023, 13, x FOR PEER REVIEW 3 of 21 Figure 1. Anatomy of the human eye. Figure 1. Anatomy of the human eye. Figure 2. Figure 2. Enlarged Enlarged optic optic cup cup within within the the optic optic disc discin inthe theglaucoma-affected glaucoma-affectedimage. image. Glaucoma Glaucoma is thethesecond secondleading leadingcause causeofof blindness blindness worldwide. worldwide. AboutAbout 80 million 80 million peo- people ple were affected by glaucoma worldwide in 2020, and the number may increase to were affected by glaucoma worldwide in 2020, and the number may increase to 111.8 111.8 million million by by 2040 2040 There There are are several several types types of of glaucoma, glaucoma, butbut the the most most common common is is open-angle open-angle glaucoma, glaucoma, which which affects affects nearly nearly 57.5 57.5 million million people people worldwide worldwide.. Regular Regular checkups checkups by Diagnostics 2023, 13, x FOR PEER REVIEW by ophthalmologists ophthalmologists afterafter age age 50 50 can can reduce reduce thethe risk risk of of developing developing glaucoma. glaucoma. 4 of 21 Figure 3 shows the retinal fundus images of a healthy control and patients Figure 3 shows the retinal fundus images of a healthy control and patients with early, with early, moderate, and advanced-stage glaucoma from the RIM-ONE moderate, and advanced-stage glaucoma from the RIM-ONE dataset. dataset. 3. The Figure 3. Figure Thefundus fundusimages imagesof of healthy control healthy (a), early control glaucoma (a), early (b), moderate glaucoma glaucoma (b), moderate (c), and (c), and glaucoma deep (advanced) deep (advanced) glaucoma glaucoma(d)(d) from the the from RIM-ONE RIM-ONE dataset. dataset. Ophthalmologists use multiple manual methods to diagnose glaucoma, including gonioscopy, pachymetry, tonometry, and perimetry. In tonometry, the IOP, a major risk factor for glaucoma, is measured. Gonioscopy measures the angle between the iris Diagnostics 2023, 13, 1738 3 of 17 Ophthalmologists use multiple manual methods to diagnose glaucoma, including gonioscopy, pachymetry, tonometry, and perimetry. In tonometry, the IOP, a major risk factor for glaucoma, is measured. Gonioscopy measures the angle between the iris and cornea. Pachymetry measures the corneal thickness. However, these manual assessment methods for glaucoma detection are very time consuming and subjective. Further, they largely depend on the availability of ophthalmologists, which can be a limiting factor in remote areas. Therefore, there is currently a need for the development of automated tools that can efficiently diagnose glaucoma at an early stage. Artificial intelligence technologies have grown significantly in recent years. Many efforts are being undertaken in healthcare to integrate AI technology for practical medical treatments [13–15]. Computer-aided diagnostic (CAD) tools for automatically detecting glaucoma are common in clinical practice. The applications of machine learning and, most recently, deep learning (DL) algorithms [16–19] have increased the diagnostic accuracy of these automated tools for detecting glaucoma. Here, an efficient and fully automated system that is based on deep learning architec- ture and can efficiently diagnose early-stage glaucoma on given datasets is proposed. The following are the main contributions of this work: The most notable recent machine learning and deep learning-based glaucoma detection research is thoroughly reviewed to define the problem, focusing on various features that can support an efficient diagnosis. For the diagnosis, a model is developed employing advanced deep learning methods along with transfer learning, and the model is tuned using various techniques to lower the likelihood of model overfitting. Multiple datasets of glaucomatous retinal images are adopted to train and test the model to achieve higher diagnostic accuracy. An end-to-end learning system that overcomes the drawbacks of current glaucoma screening methods is developed. The remaining part of the paper is organized into the following themes: The previous work by other researchers is explained in Section 2. Section 3 explains the proposed method- ology. Section 4 describes the experimentation and results of the proposed model. Section 5 presents a discussion of the results, and Section 6 presents a conclusion summarizing the key findings. 2. Literature Review Researchers have developed several techniques for the detection of glaucoma. Among these techniques, machine learning-based methods manually extract the features and perform classification by using different machine learning classifiers. Recently, deep learning models, such as convolutional neural networks (CNNs), have been widely used to diagnose diseases automatically without human involvement. Glaucoma detection through CNNs is performed by various researchers [21–30]. The CNN-based systems perform effective computation and provide robust results for disease classification. A CNN consists of different layers, such as convolutional, activation, pooling, and the fully connected layer (FCL). Each architecture consists of a different combination of these layers. The diagnosis and detection of other retinal diseases such as papilledema [23,31], diabetic retinopathy , central serous retinopathy (CSR) [32,33], and hypertensive retinopathy can be performed through deep learning and machine learning methodologies using OCT and fundus images [5,30]. Diabetes and other eye diseases have been successfully diagnosed by DL techniques. The application of the CAD system has widened the diagnostic horizon in several other disease diagnoses, such as CSR , lung tumor , brain tumor , skin tumor , and prostate cancer. The fundus images provide a clear picture of the eye’s internal structure and are widely used for glaucoma diagnosis. The glaucoma classification using fundus images through DL models has shown encouraging results [36,37]. The fundus images clearly depict the optic nerve head and are readily available for training the glaucoma Diagnostics 2023, 13, 1738 4 of 17 detection models. Various models based on pre-trained CNN models [14,39], ensemble approaches [40–42] and CNN-based architectures are encountered in this article for the detection of glaucoma. Serte and Serener developed a glaucoma detection model using an ensemble approach based on a local dataset of 1542 fundus images. The model cropped the OD by using a graph saliency region technique. Three CNN architectures, namely ResNet-50, ResNet-152, and AlexNet, were used as the ensemble classifiers in this model. All three methods, including without saliency map, with saliency map and single CNN model, and with saliency map and ensemble approach, were tested, and the best results were obtained for the ensemble approach with an AUC of 94% and accuracy of 88%. Chaudhary and Pachori developed a glaucoma detection model based on two methods, using RIM-ONE, ORIGA, and DRISHTI-GS datasets. The 2D Fourier–Bessel series expansion-based empirical wavelet transform was used for the segmentation of the boundary. Two methods were used, one depending on the ML model and the other using the ensemble approach of the CNN architecture ResNet. The first model at full scale obtained the best results. The best results with the second method were obtained with the ensemble technique at a full scale with 91.1% accuracy, 91.1% sensitivity, 94.3% specificity, 83.3% AUC, and 96% ROC. GlaucomaNet was proposed to identify POAG based on dataset images from different populations and settings. The model comprises two CNNs intended to mimic the human grading process. To this end, the first CNN learns the discriminative features, whereas the second fuses the features for grading. This simulation of the human grading process combined with an ensemble of network architectures greatly enhanced the diagnostic accuracy. Thakoor et al. developed a model based on different CNN architectures trained on OCT images and also used some pre-trained models to detect glaucoma. The pre- trained ResNet, VGG, and InceptionNet were combined with random forest and compared with the CNN architectures trained on OCT images. A high accuracy of 96.27% was achieved with the CNN trained on the OCT images. Hemelings et al. proposed an approach for glaucoma detection using pre-trained ResNet-128 architecture with 7083 OD center fundus images. The transfer and active learning approaches were used to enhance the diagnosis capability of the model. The use of a saliency map highlighted the affected region to provide evidence of the disease. The model achieved robust results with an AUC of 99.55%, a specificity of 93%, and a sensitivity of 99.2% for glaucoma detection. Yu et al. developed a model using a modified version of U-Net architecture in fundal images for glaucoma diagnosis using multiple datasets. The U-Net used the pre- trained ResNet-34 as an encoder and the classical U-Net architecture as a decoder. The model showed good performance as 97.38% of disc dice values and 88.77% of cup dice values were aligned with the DRISHTI-GS test set. Other authors proposed an approach named AG-CNN, which detected glaucoma and localization of pathological areas using the fundus images. The model is based on attention prediction, localization of the affected area, and glaucoma classification. The deep features predicted glaucoma through the visual maps of necrotic areas in the LAG and RIM-ONE datasets. The use of attention maps for localizing the pathological area demonstrated high efficacy. The model prediction for glaucoma was superior to previous models, with an accuracy of 95.3%. Phan et al. developed a model based on three CNN architectures, ResNet-152, VGG19, and DenseNet201, for diagnosing glaucoma on 3312 retinal fundus images. The proposed model has also been tested on poor-quality images to examine its diagnostic accuracy in glaucoma. All the architectures achieved an AUC of 90% for detecting glaucoma. Liao et al. proposed a novel CNN-based scheme that used ResBlock architecture to diagnose glaucoma using the ORIGA dataset. The model diagnosed glaucoma and provided a transparent interpretation based on visual evidence by highlighting the affected area. The model named EAMNet contained three parts: ResNet architecture extracted the features and aggregation, and the multiple-layer average pooling (M-LAP) linked the semantic detail and information of the localization, while the evidence activation map Diagnostics 2023, 13, 1738 5 of 17 (EAP) was used for the evidence of the affected area the physician used for the final decision. The activation map was used to provide the clinical basis for glaucoma. The proposed scheme efficiently diagnosed glaucoma, with an AUC of 0.88. Researchers developed the G-Net model based on CNN to detect glaucoma in the DRISHTI-GS dataset. The model used two neural networks (U-Net) to separate the disc and cup. The cropped fundus images in the red channel were fed to the model. The model contained 31 layers of convolutional, max-pooling, up-sampling, and merge layers. The filters applied were of sizes (3, 3), (1, 1), and (1, 32), and 64 filters were used on different layers. The model labeled the pixel as black on segmenting the OD in the real image and white otherwise. The output images were fed to the other model to segment the cup. The second model was like the first model, with a single difference in the size of the filters (4, 4). The output of this model was a segmented cup. These two outputs were used to calculate the CDR for the glaucoma prediction. This algorithm used two neural networks to obtain a high accuracy of 95.8% for OD and 93.0% for OC segmentation. Researchers developed a model based on CNN for glaucoma detection using 1110 OCT images and compared its performance with the ML algorithms. A total of 22 features were extracted and fed to different machine learning classifiers such as NB, RF, SVM, LR, Gradient Adaboost, and Extra Trees. The CNN model classified and achieved better results with an AUC of 0.97 than other machine learning approaches, such as logistic regression, with an AUC of 0.89. Thakur et al. proposed a model capable of diagnosing glaucoma before the onset of the disease. Three deep learning models were trained on 66,721 fundus images that can detect glaucoma, such as 1 to 3 years ago, 4 to 7 years ago, and before the onset of glaucoma. All three models achieved AUCs of 0.88, 0.77, and 0.97 in detecting glaucoma. Lima et al. developed a CNN model for the optic cup segmentation for the detection of glaucoma. The modified U-Net architecture segmented the optic cup from the green channel image, and the optic disc mask was given as input. The model achieved a dice value of 94% on the DRISHTI dataset. Maheshwari et al. presented a model that converted the images into RGB channels after dividing the dataset images into training and testing images. The LBP-based augmentation was applied to obtain the best results. The model achieved 98.90% accuracy, 100% sensitivity, and 97.50% specificity. Lima et al. used a genetic model based on CNN with 25 layers using the RIM-ONE dataset to diagnose glaucoma. The model achieved an accuracy of 91% in detecting glaucoma. Saxena et al. developed a six-layer CNN model for glaucoma detection using the SCES and the ORIGA datasets. The ROI was extracted using the ARGALI approach, and the data augmentation technique was used to avoid the overfitting problem. The model achieved excellent results, with an AUC of 0.882 on SCES and 0.822 on ORIGA datasets. Elangovan and Nath developed a CNN-based model consisting of 18 layers for glaucoma detection. The model was based on DRISHTI– GS1, ORIGA, RIM–ONE2 (release 2), ACRIMA, and LAG datasets. The best results were obtained with the ACRIMA dataset, achieving 96.64% accuracy, 96.07% sensitivity, 97.39% specificity, and 97.74% precision. Aamir et al. developed a multi-level CNN model for diagnosing glaucoma. The fundus images were preprocessed to reduce noise with the adaptive histogram equalizer technique. The model classified the fundus images for glaucoma detection into advanced, moderate, and early categories. The model achieved a sensitivity of 97.04%, a specificity of 98.99%, an accuracy of 99.39%, and a PRC of 98.2% on 1338 fundus images. Raja et al. proposed a technique for diagnosing glaucoma using a dataset of 196 OCT images. The proposed model used CNN and calculated the CDR with 94% accuracy, 94.4% sensitivity, and 93.75% specificity in detecting glaucoma. Carvalho et al. proposed a 3DCNN algorithm for diagnosing glaucoma through the fundus images of RIM-ONE and DRISHTI-GS datasets. The 2D fundus images were converted into 3D volumes for each RGB and gray channel. The CNN was trained on all four channels and showed the best results on a gray channel with 83.23% accuracy, 85.54% sensitivity, 80.95% specificity, 83.2% AUC, and 66.45 Kappa. Diagnostics 2023, 13, 1738 6 of 17 Gheisari et al. developed a combined model based on a CNN and a recurrent neural network for diagnosing glaucoma using retinal fundus images. The diagnostic results were achieved with an F-measure of 96.2% on 295 videos and 1810 fundus images. Veena et al. developed a CNN model for the detection of glaucoma. The images were preprocessed to eliminate the noise using the Gaussian filter. The Sobel edge and the watershed algorithms extracted the features from the fundus images. The model achieved the OD and OC segmentation accuracies of 0.9845 and 0.9732, respectively, on the DRISHTI dataset. The achieved results are 98.48% accuracy, 99.3% sensitivity, 96.52% specificity, 97% AUC, and 98% of F1-score on the G1020 dataset. Recently, Fan et al. assessed the diagnostic precision, generalizability, and ex- plainability of a Vision Transformer deep learning method in diagnosing the primary open-angle glaucoma and identifying the salient areas found in the retinal images. A dual learning-based technique that combines deep learning and machine learning was proposed by Thanki. For identifying distinctive retinal characteristics, a deep neural network extracts deep features. Following that, a hybrid classification algorithm is employed to accurately classify glaucomatous retinal images. The following Table 1 shows the summary of year-wise published studies for the detection of glaucoma. Table 1. Summary of Literature Review. Sr. No. Authors Year Model Datasets Results Pre-trained RIGA, Dice 97.38% (Disc) 1 Yu et al. 2019 U-Net, DRISHTI-GS, Dice 88.77% (Cup) ResNet RIM-ONE LAG, 2 Li et al. 2019 CNN Accuracy 95.3% RIM-ONE ResNet-152, Local dataset 3 Phan et al. 2019 DenseNet201, AUC 0.9 of 3777 images VGG19 4 Liao et al. 2019 ResNet ORIGA Accuracy 0.88 ResNet-50, HRF, ResNet-152, DRISHTI-GS1, Accuracy 53%, and 5 Serte et al. 2019 RIMONE, AUC 83%, GoogleNet sjchoi86-HRF, specificity 100% (ensemble ACRIMA method) Accuracy 95.8% (OD segmentation), 6 Juneja et al. 2019 U-Net DRISHTI-GS 93.0% (OC segmentation) Local dataset 7 Maetschke et al. 2019 CNN AUC 0.94 of 1110 images Pre-trained Local dataset 8 Thakoor et al. 2019 Accuracy 96.27% CNN of 737 images Accuracy: 98.90% 9 Maheshwari et al. 2020 AlexNet RIM-ONE Sensitivity: 100% Specificity: 97.50% 10 Lima et al. 2020 CNN RIM-ONE r3 Accuracy 91% AUC 0.822 11 Saxena et al. 2020 CNN ORIGA, SCES (ORIGA) AUC 0.882 (SCES) Local datasets MobileNet of 45,301, 12 Thakur et al. 2020 AUC 0.97 v2 42,601, and 42,498 images AUC 0.995 Pre-trained Local dataset 13 Hemelings et al. 2020 Sensitivity 99.2% ResNet 128 of 1424 images Specificity 93% Diagnostics 2023, 13, 1738 7 of 17 Table 1. Cont. Sr. No. Authors Year Model Datasets Results RIM-ONE, Accuracy 96.64%, DRISHTI–GS1, sensitivity 96.07%, 14 Elangovan and Nath 2020 CNN ORIGA, LAG, specificity 97.39%, ACRIMA precision 97.74% Sensitivity 97.04%, Local dataset specificity 98.99%, 15 Aamir et al. 2020 ML-DCNN of 1338 fundus accuracy 99.39%, images PRC 98.2% Local dataset Accuracy 94%, 16 Raja et al. 2020 CNN of 196 OCT sensitivity 94.4%, images specificity 93.75% 295 videos and local dataset of 17 Gheisari et al. 2021 CNN, RNN F-measure 96.2% 1810 fundus images Accuracy 91.1%, Ensemble RIM-ONE, sensitivity 91.1%, 18 Chaudhary and Pachori 2021 ResNet ORIGA, and specificity 94.3%, Models DRISHTI-GS AUC 83.3%, ROC 96% Accuracy 83.23%, sensitivity 85.54%, RIM-ONE and 19 Carvalho et al. 2021 3DCNN specificity 80.95%, DRISHTI-GS AUC 83.2%, and Kappa 66.45% Accuracy 0.930 OHTS and 20 Lin et al. 2022 CNN (OHTS) and 0.969 LAG (LAG) Accuracy 98% 21 Veena et al. 2022 CNN DRISHTI-GS (OD), 97% (OC) Custom assembled 22 Fan et al. 2023 CNN AUC 0.91 from 5 public datasets DRISHTI-GS 23 Thanki 2023 Deep NN Accuracy 100% and ORIGA 3. Proposed Methodology The innovation in artificial intelligence may help in a fast and accurate diagnosis of diseases. The proposed model is developed using the ResNet-50 robust image classifica- tion architecture. The fundus imaging modality is used as it precisely depicts the eye’s internal structure. The applications of fundus images are numerous for many other disease diagnoses, such as cataracts, retinopathy of prematurity, DR, and age-related macular degeneration (AMD). Figure 4 depicts the flow diagram of the proposed model. 3.1. Dataset In this research, four publicly available datasets are used for testing and training the model: (i) G1020 , (ii) DRISHTI-GS , (iii) RIM-ONE , and (iv) ORIGA. The G1020 dataset comprises 1020 fundus images with high resolution, CDR calculation, OD and OC segmentation, size of the neuro-retinal rim in inferior, superior, nasal, and temporal regions, and location of the bounding box for OD for glaucoma detection. The images in the dataset are only focused on the fundus region by removing the unrelated region. The size of the images is between 1944 × 2108 and 2426 × 3007 pixels. The dataset is publicly available.  3. Proposed Methodology The innovation in artificial intelligence may help in a fast and accurate diagnosis of diseases. The proposed model is developed using the ResNet-50 robust image classifica- tion architecture. The fundus imaging modality is used as it precisely depicts the eye’s internal structure. The applications of fundus images are numerous for many other dis- Diagnostics 2023, 13, 1738 8 of 17 ease diagnoses, such as cataracts, retinopathy of prematurity, DR, and age-related macu- lar degeneration (AMD). Figure 4 depicts the flow diagram of the proposed model. Figure4. Figure Flowdiagram 4.Flow diagramofofthe theproposed proposedmodel modelfor forglaucoma glaucomadetection. detection. The DRISHTI-GS dataset contains OD and OC segmented and ground truth images. 3.1. Dataset This dataset contains 101 images, of which 31 are healthy and 70 include eyes with glaucoma. In this research, four publicly available datasets are used for testing and training the The fundus images in this dataset are focused on the OD with a field of view of 30 degrees, model: (i) G1020 , (ii) DRISHTI-GS , (iii) RIM-ONE , and (iv) ORIGA. The and the image resolution is 2896 × 1944 pixels in PNG format. Six experts performed the G1020 dataset comprises 1020 fundus images with high resolution, CDR calculation, OD manual annotation of OD and OC in this dataset. This dataset is publicly available. The and OC segmentation, size of the neuro-retinal rim in inferior, superior, nasal, and tem- RIM-ONE dataset comprises 169 ONH segmented high-resolution fundus images. The poral regions, and location of the bounding box for OD for glaucoma detection. The im- images are captured with a fundus camera (Nidek AFC-210). There are four categories of ages in the dataset are only focused on the fundus region by removing the unrelated re- images, including 118 normal, 12 early glaucoma, 14 moderate, 14 deep, and Diagnostics 2023, 13, x FOR PEER REVIEW 10 of 11 21 images gion. The size of the images is between 1944 × 2108 and 2426 × 3007 pixels. The dataset is for ocular hypertension. These images are also publicly available. The ORIGA dataset publicly available. comprises 650 segmented and annotated images. Every image is labeled with grading The DRISHTI-GS dataset contains OD and OC segmented and ground truth images. information. dataset comprises This dataset can be used 650 segmented for imageimages. and annotated processingEveryalgorithms and thewith image is labeled method for This dataset contains This grading information. 101 images,can of which 31forareimage healthy and 70 algorithms include eyes with glau- detection of peripapillarydataset be used atrophy (PPA) detection andprocessing the junction of the and disctheboundary coma. method The forfundus detectionimages in this dataset of peripapillary are(PPA) atrophy focused on theand detection ODthe with a field junction of of theview disc of 30 blood vessels. degrees, boundary and the vessels. blood image resolution is 2896 × 1944 pixels in PNG format. Six experts per- formed the manual 3.2. Image Preprocessing annotation of OD and OC in this dataset. This dataset is publicly avail- able. The RIM-ONE 3.2. Image Preprocessing dataset comprises 169 ONH segmented high-resolution fundus im- Image preprocessing is performed to expand the image quality for further analysis. ages. The images Image preprocessingare captured with atofundus is performed expand camera (Nidek AFC-210). for furtherThere are four This often helps to produce more robust resultsthefromimage the quality CNN architecture. analysis. The greyscale categories This oftenof images, helps including to produce more118 normal, robust results12from earlytheglaucoma, 14 moderate, CNN architecture. 14 deep, and The greyscale images are obtained from all the training images collected from the G1020, DRISHTI- 11images images arefor ocular from obtained hypertension. These all the training images images are also collected frompublicly available. the G1020, The ORIGA DRISHTI-GS, GS, ORIGA, and RIM-ONE datasets. The grayscale image modality provides a clear and ORIGA, and RIM-ONE datasets. The grayscale image modality provides a clear and sharper sharperview view of the fundus of the fundusimages, images,asasdisplayed displayed in in Figure Figure 5. 5. TheThe grayscale grayscale morphology morphol- synthesizes ogy synthesizes all pixels with all pixels withaahomogeneous intensity homogeneous intensity value. value. Alltraining All the the training images images are are converted converted into graychannels. into gray channels.TheThe OD-centered OD-centered images images in grayscale in grayscale aretofed are fed thetoResNet- the ResNet-50 model 50 modelforfortraining. training. (a) (b) Figure5. Figure 5. Fundus Fundus image image with withgrayscale grayscaleconversion conversion(a) (a)Normal Normalfundus image fundus (b)(b) image Greyscale fundus Greyscale fundus image. image 3.3. Data Augmentation The data augmentation technique has been used to increase the number of images when the available data are inadequate for statistical and biological significance. The aug- Diagnostics 2023, 13, 1738 9 of 17 3.3. Data Augmentation The data augmentation technique has been used to increase the number of images when the available data are inadequate for statistical and biological significance. The augmentation technique is a better approach to overcome this problem due to the limited availability of images in the medical field. This technique slightly modifies the existing data to create more copies of the data. The data augmentation technique can also over- come the overfitting of the deep learning models by enhancing the model’s performance and diagnostic capability. Different techniques are applied, such as flipping the images horizontally and vertically, rotation, cropping, and scaling. 3.4. Transfer Learning The deep learning model’s training from scratch is tedious work requiring a large image dataset and efficient hardware. Additionally, it also requires more training time. The transfer learning approach uses the pre-trained model, which is trained on a large number of images such as ImageNet. It transfers the knowledge learned from the model to another model even if the field is different. The pre-trained model is trained according to the new data by changing some parameters. In this work, the pre-trained CNN architecture ResNet-50 is retrained on the G1020, ORIGA, RIM-ONE, and DRISHTI-GS datasets. 3.5. Convolutional Neural Network The CNN is a multilayer DL network that obtains the input as high-dimension data (images) and progressively extracts high-dimension features from the input images. The CNN architectures consist of different numbers of layers, which increase as the size of the input images increases. The network learns more accurately as it goes deeper. However, the major drawback of the deeper networks is the increase in computation time. CNNs have shown promising image processing, object detection, image segmentation, image classification, video processing, and natural language processing features. The appli- cations of CNN architectures have shown tremendous results for disease diagnosis in the medical sciences. 3.6. ResNet-50 Architecture The ResNet is the short form of the residual network, and it solves the vanishing gra- dient problem by using the skip connection approach. Before ResNet, network degradation problems occurred due to increased network depth. The result of this degradation was a higher training error. To overcome this problem, the skip connection technique is applied in the ResNet architecture. This architecture shows higher detection accuracy, takes less training time, and is easier to optimize. The ResNet architecture has several applications for image processing and diagnosis of diseases in the medical field. Additionally, it has shown excellent results for object detection and face recognition. Figure 6 shows the architecture of ResNet-50. The difference in the ResNet-50 from the earlier ResNet-18 and ResNet-34 is skipping three layers instead of two and using a 1 × 1 convolution layer. There are 50 ayers in this architecture, and it is capable of classifying data into seven classes. It is widely applied for image recognition, object localization, and object detection. Consequently, it has considerably reduced computational costs. The block diagram of the proposed methodology for glaucoma detection is shown in Figure 4. The proposed methodology consists of the following steps: Acquire the fundus images from different publicly available datasets. Convert the fundus images into grayscale. Apply the data augmentation approach to multiply the number of images by flipping, rescaling, and rotation after dividing the dataset into training and testing sets. Further, 80% of the images in the dataset are used for training, 10% of images for validation, and the remaining 10% for testing. Pre-trained DL architecture, such as the ResNet-50, is used for classification. The model classifies an image as either a healthy or glaucomatous image.  plications for image processing and diagnosis of diseases in the medical field. Addition- ally, it has shown excellent results for object detection and face recognition. Figure 6 shows the architecture of ResNet-50. The difference in the ResNet-50 from the earlier ResNet-18 and ResNet-34 is skipping three layers instead of two and using a 1 × 1 convo- lution layer. There are 50 layers in this architecture, and it is capable of classifying data Diagnostics 2023, 13, 1738 10 of 17 into seven classes. It is widely applied for image recognition, object localization, and object detection. Consequently, it has considerably reduced computational costs. Figure 6. Block Figure 6. Block diagram diagram of of ResNet-50 architecture. ResNet-50 architecture. 4. Experiments and Results The block diagram of the proposed methodology for glaucoma detection is shown in The Figure 4. proposed model The proposed is evaluated methodology using performance consists metrics of the following steps:such as accuracy, sen- sitivity, and specificity. There are four possibilities for the classified images, namely true Acquire the fundus images from different publicly available datasets. positive, true negative, false positive, and false negative. The true positive labels the image Convert the fundus images into grayscale. as affected by glaucoma, and it is a correct prediction. The true negative labels the image Apply the data augmentation approach to multiply the number of images by flip- as a healthy image, and it is also correctly classified. The false positive erroneously labels ping, rescaling, and rotation after dividing the dataset into training and testing sets. an image as a glaucoma-affected, otherwise healthy image. The false negative incorrectly Further, 80% of the images in the dataset are used for training, 10% of images for labels a glaucoma-affected image as a healthy image. validation, and the remaining 10% for testing. The accuracy is the measure of the correctly labeled images divided by the total Pre-trained DL architecture, such as the ResNet-50, is used for classification. number of images. It can be calculated as in Equation (1). The model classifies an image as either a healthy or glaucomatous image. TP + TN 4. Experiments and ResultsAccuracy = TP + TN + FP + FN (1) The proposed model is evaluated using performance metrics such as accuracy, sen- Theand sitivity, sensitivity represents specificity. thefour There are correctly classified possibilities images for the affected classified by glaucoma. images, It is namely true calculated as in Equation (2): positive, true negative, false positive, and false negative. The true positive labels the image as affected by glaucoma, and it is a correct prediction. TP The true negative labels the image as a healthy image, and it is also correctly = Senstivityclassified. (2) The false positive erroneously labels TP + FN an image as a glaucoma-affected, otherwise healthy image. The false negative incorrectly labels Thea glaucoma-affected image specificity represents theas a healthy correctly image. healthy images. It can be calculated classified as Equation (3): TN Speci f icity = (3) TN + FP The F1-score can be calculated as in Equation (4): 2TP F1 − Score = (4) 2TP + FP + FN The dataset’s fundus images were divided into three subcategories: training, vali- dation, and testing. We used 80% of the images for the training of the model, 10% for model validation, and the remaining 10% for testing. All the images were resized to the same size and centered on the optic disc. Moreover, the model was trained using the SDG solver with a learning rate of 0.001 on ten epochs in Python with a system configu- ration of Intel/Xeon/CPU E3-1225, 3.3 GHz, and 16 GB RAM. The computational time for training of model on these datasets was 30 min. Figure 7 shows the number of images before and after data augmentation. Four datasets, namely G1020, RIM-ONE, ORIGA, and DRISHTI-GS, were used. The ResNet-50 architecture achieved robust results with 98.48% accuracy, 96.52% specificity, 99.30% sensitivity, 97% AUC, and an F1-score of 98% on the G1020 dataset. The comparison of the proposed model with the previous studies is shown in Table 2. Figures 8–10 show the accuracy and error rate for the training data using ResNet-50 over G1020, DRISHTI-GS, RIM-ONE, and ORIGA datasets, respectively. Figure 11 shows the confusion matrix of validation data of all four datasets.  Diagnostics 2023, 13, 1738 11 of 17 Table 2. Comparison Table of the Proposed Model with State-of-the-Art Models. Sr # Authors Dataset AUC Accuracy Sensitivity Specificity F1-Score 1 Lima et al. RIM-ONE r3 91% - - - - 2 Diagnostics 2023, 13, x FOR PEER REVIEW Saxena et al. SCES 88.2% - - - - 13 Local dataset 3 Thakoor et al. - 96.27% - - - of 737 images OHTS 91% DIGS 74% Veena Fan et al. et al. - - - - DRISHTI–GS ACRIMA 74% 98% 95.41 LAG 79% Gomez- RIM-ONE 90% Local dataset of 4 Valverde et ORIGA 94%55% 87.01% 89.01% 89.01% - 2313 images OHTS Linal. et al. LAG 90.4% 93% 49% Christopher Thanki Local dataset of ORIGA 69.7% 76.2% 100% 73% 5 Veena et al. DRISHTI–GS 97% 98% 88% 95% 95% 95.41% - et al. 14,822 images Local dataset of 4 Gomez-Valverde et al. Local datasets of 94% 87.01% 89.01% 89.01% - 2313 images 5 Thakuret et Christopher al. 45,301, al. 42,601, Local dataset of 97% 6 14,822 images 97%88% - 95% -95% -- - and 42,498 Local datasets of im- 45,301, 42,601, 6 Thakur et al. ages and 42,498 97% - - - - images RIM-ONE 94.2% 96.15% 97.85% 92.38% 97% RIM-ONE 94.2% 96.15% 97.85% 92.38% 97% Proposed ORIGA ORIGA 93% 93%92.59% 92.59% 98.39% 98.39% 79.26% 79.26% 95% 95% Proposed Method Method G1020 G1020 97% 97% 98.48% 98.48% 99.30% 99.30% 96.52% 96.52% 98% 98% DRISHTI-GS DRISHTI-GS 96% 96% 97.03% 97.03% 93.75% 93.75% 98.55% 98.55% 97% 97% DRISHTI-GS ORIGA RIM-ONE G1020 310 After Augmentation Healthy 4820 1180 7240 700 Glaucoma 1680 510 2960 31 Before Augmentation Healthy 482 118 724 70 Glaucoma 168 51 296 Figure7.7.Dataset Figure Datasetcomparison before comparison and after before anddata augmentation. after data augmentation. Diagnostics 2023, Diagnostics 13,13,x 1738 2023, FOR PEER REVIEW 12 of 17 14 Diagnostics 2023, 13, x FOR PEER REVIEW 15 Figure8.8.Training Figure Training process process curvecurve of G1020 of G1020 dataset.dataset. Figure9. 9. Figure Training Training process process curvecurve of the of the DRISHTI-GS DRISHTI-GS dataset. dataset. Diagnostics Diagnostics2023, 2023,13, 13,x1738 FOR PEER REVIEW 13 of 17 16 of 2 Diagnostics 2023, 13, x FOR PEER REVIEW 17 of 21 Figure10. Figure 10.Training Training process process curve curve of RIM-ONE of RIM-ONE dataset. dataset. Figure11. Figure Confusion 11.Confusion matrix matrix forfor data data validation validation on on (a) (a) DRISHTI-GS, DRISHTI-GS, (b) (b) G1020, G1020, (c) ORIGA, (c) ORIGA, and and (d) (d) RIM-ONE RIM-ONE datasets. datasets. 5. Discussion 5. Discussion The proposed model uses the deep learning architecture ResNet-50 to diagnose early- The proposed model uses the deep learning architecture ResNet-50 to diagnose early- stage glaucoma using fundus images. Four datasets, G1020, DRISHTI-GS, RIM-ONE, stage glaucoma using fundus images. Four datasets, G1020, DRISHTI-GS, RIM-ONE, and and ORIGA, were used for the proposed model’s training, validation, and testing. The ORIGA, were used for the proposed model’s training, validation, and testing. The capa- capability of deep learning models for automatic identification of the pattern from images bility of deep learning models for automatic identification of the pattern from images has has smoothed the data for obtaining robust results for disease detection. The greater smoothed the data for obtaining robust results for disease detection. The greater number number of layers in the model requires more training time, and sometimes deeper models of layers in the model requires more training time, and sometimes deeper models take take several weeks for training, which is not optimal in clinical settings. The pre-trained several weeks for training, which is not optimal in clinical settings. The pre-trained Res- ResNet-50 architecture can best classify the images in reduced computation time. The Net-50 architecture can best classify the images in reduced computation time. The training training of the model from scratch requires a large amount of data and training time. So, of the the model learning transfer from scratch requires approach is aapplied large amount to saveofcomputation data and training time.achieve time and So, the robust trans- fer learning approach is applied to save computation time and achieve robust diagnostic diagnostic results. The use of pre-trained models while training the CNN architectures for results. Thehas a new task use made of pre-trained models it possible whileatraining to develop fast andthe CNNdiagnosis reliable architectures for adespite system, new task the has made it possible to develop a fast and reliable diagnosis system, despite the limited availability of the required data. The dataset’s limited images cause overfitting limited of availability of the required data. The dataset’s limited images cause overfitting of the model, but the data augmentation technique overcomes this problem. Many images can be created from a single image, providing a large dataset for the DL models for training. The fundus images are a cheap solution for the diagnosis of glaucoma. The fundus images in the gray channel depict the lesion more precisely and clearly indicate the af- Diagnostics 2023, 13, 1738 14 of 17 the model, but the data augmentation technique overcomes this problem. Many images can be created from a single image, providing a large dataset for the DL models for training. The fundus images are a cheap solution for the diagnosis of glaucoma. The fundus images in the gray channel depict the lesion more precisely and clearly indicate the affected region. The G1020, RIM-ONE, ORIGA, and DRISHTI-GS datasets, which contain the OD segmented images, were applied. The proposed model has exhibited glaucoma detection with 98.48% accuracy, 99.30% sensitivity, 96.52% specificity, an AUC of 97%, and an F1-score of 98% on the G1020 dataset. The proposed model’s results on the ORIGA dataset include 92.59% accuracy, 98.39% sensitivity, 79.26% specificity, 93% AUC, and 95% F1-score. The RIM-ONE dataset has shown 96.15% accuracy, 97.85% sensitivity, 92.38% specificity, 94.2% AUC, and 97% F1-score on the proposed model. The DRISHTI-GS has shown 97.03% accuracy, 93.75% sensitivity, 98.55% specificity, 96% AUC, and 97% F1-score. The results of the proposed model on all four datasets are shown in Table 2. The proposed model has shown more robust results than the existing techniques on the G1020 dataset. Due to the wide availability of high-resolution images in the G1020 dataset, the best performance of the proposed model is obtained on the G1020 dataset. The performance of the proposed model is poor on the ORIGA dataset compared to other datasets in terms of specificity. This