Untitled

• Glaucoma is characterized by increased intraocular pressure and optic nerve damage, leading to irreversible blindness.
• Early-stage detection is crucial to avoid the drastic effects of glaucoma.
• Glaucoma is frequently detected at an advanced stage in the elderly.
• Manual assessment methods are costly, time-consuming, and require skilled ophthalmologists.
• There is no definitive diagnostic technique for early-stage glaucoma.
• An automatic deep learning method is presented for detecting early-stage glaucoma with high accuracy.
• The technique identifies patterns in retinal images often overlooked by clinicians.
• The method uses gray channels of fundus images and data augmentation to train a convolutional neural network model.
• The ResNet-50 architecture is used.
• Datasets used: G1020, RIM-ONE, ORIGA, and DRISHTI-GS.
• The proposed model helps clinicians diagnose early-stage glaucoma for timely interventions.
• Keywords: glaucoma, fundus images, deep learning, early-stage detection, augmentation.

Introduction

• Major eye components for vision: cornea, pupil, iris, lens, retina, optic nerve, and tears.
• The iris controls the amount of light entering the eye.
• The retina converts light into electrical signals, which are sent to the brain.
• The optic nerve transmits visual signals (1 million nerve fibers) from the retina to the occipital cortex.
• Aqueous humor is a fluid in the eye that is continuously recycled.
• Obstruction of aqueous humor drainage increases intraocular pressure (IOP).
• Increased IOP can damage the retina and optic nerve, possibly leading to vision loss.
• Damage is partly due to the degeneration of ganglion cells in the retina.
• Loss of nerve fibers alters the shape of the optic disc, increasing the cup-to-disc ratio (CDR), which is an early sign of glaucoma.
• Visual loss results from damage to retinal ganglionic cells.
• Alterations in the visual field scope are essential for diagnosing glaucoma.

Glaucoma Statistics

• Glaucoma is the second leading cause of blindness worldwide.
• About 80 million individuals globally were affected by glaucoma in 2020.
• Estimates suggest that this figure could reach 111.8 million by 2040.
• Open-angle glaucoma is the most common type, affecting around 57.5 million people globally.
• Regular checkups by ophthalmologists aged 50 and above can decrease the risk of glaucoma development.

Glaucoma Diagnosis Methods

• Multiple manual methods used to diagnose glaucoma include gonioscopy, pachymetry, tonometry, and perimetry.
• Tonometry measures IOP.
• Gonioscopy measures the angle between the iris and cornea.
• Pachymetry measures corneal thickness.
• Manual assessment methods are time-consuming and subjective.
• Availability of ophthalmologists is a limiting factor in remote areas.
• Automated tools are needed that can efficiently diagnose glaucoma early.

Advancements in AI for Glaucoma Detection

• Artificial intelligence technologies have grown significantly.
• AI technology is being integrated into healthcare for practical medical treatments.
• Computer-aided diagnostic (CAD) tools automatically detect glaucoma in clinical practice.
• Machine and deep learning algorithms have increased the diagnostic accuracy of automated tools for detecting glaucoma.

Proposed Deep Learning System

• Proposes an efficient, automated system based on deep learning architecture for early-stage glaucoma diagnosis using given datasets.
• Reviews recent machine learning and deep learning-based glaucoma detection research focusing on features for efficient diagnosis.
• Employs advanced deep learning methods and transfer learning, tuning the model to reduce overfitting likelihood.
• Adopts multiple glaucomatous retinal image datasets to train/test the model to achieve higher diagnostic accuracy.
• Develops an end-to-end learning system that overcomes current glaucoma screening methods' drawbacks.

Literature Review Summary

• Researchers have developed machine learning-based methods and deep learning models like CNNs for glaucoma detection.
• CNNs perform computation effectively and give robust results for disease classification using different layers like convolutional, activation, pooling, and FCL.
• Deep learning and machine learning can perform diagnosis and detection of other retinal diseases (papilledema, diabetic retinopathy, CSR, and hypertensive retinopathy) through OCT and fundus images.
• CAD systems have widened the diagnostic horizon in other diseases like CSR, lung tumor, brain tumor, skin tumor, and prostate cancer.
• Fundus images provide a clear picture of the eye's internal structure, which is used for glaucoma diagnosis through deep learning models.

Related Glaucoma Detection Models

• Serte and Serener developed a glaucoma detection model using a local dataset of 1542 fundus images and an ensemble approach with three CNN architectures (ResNet-50, ResNet-152, and AlexNet). The ensemble approach achieved an AUC of 94% and accuracy of 88%.
• Chaudhary and Pachori developed a glaucoma detection model using RIM-ONE, ORIGA, and DRISHTI-GS datasets, which consist of a ML model and an CNN architecture (ResNet) ensemble approach.
• GlaucomaNet was proposed to identify POAG based on images from different populations and used two CNNs intended to mimic the human grading process.
• Thakoor et al. created a model based on CNN architectures trained on OCT images and pre-trained models to detect glaucoma with a high accuracy, and Hemelings et al. used pre-trained ResNet-128 architecture.
• Yu et al. developed a model using a modified version of U-Net architecture in fundal images achieving good performance.
• Phan et al. developed a model based on three CNN architectures and achieved an AUC of 90% for detecting glaucoma.
• Liao et al. proposed a CNN-based scheme using ResBlock architecture and the model named EAMNet.
• Researchers developed the G-Net model and used two neural networks (U-Net) to separate the disc and cup in DRISHTI-GS dataset, achieving a high accuracy.
• Researchers created a CNN-based model for glaucoma detection using 110 OCT images.
• Lima et al. created a CNN model for the optic cup segmentation for the detection of glaucoma.
• Saxena et al. developed a six-layer CNN model for glaucoma detection.
• Carvalho et al. proposed a 3DCNN algorithm for diagnosing glaucoma through fundus images.

Proposed Methodology

• A model is developed using ResNet-50, a reliable image classification architecture.
• Fundus imaging modality is employed due to its precision in depicting the eye's internal structure.
• Numerous applications of fundus images, including diagnoses of cataracts, retinopathy of prematurity, DR, and age-related macular degeneration (AMD).

Datasets Used

• G1020: High-resolution fundus images focused on the fundus region.
• DRISHTI-GS: OD and OC segmented and ground truth images focused on the OD with manual annotation.
• RIM-ONE: ONH segmented high-resolution fundus images with images captured by fundus camera.
• ORIGA: Segmented and annotated images labeled with grading information used for PPA detection and disc boundary.

Image Preprocessing

• Image preprocessing is performed to enhance image quality for analysis.
• Grayscale images are derived from all training images from the G1020, DRISHTI-GS, ORIGA, and RIM-ONE datasets.
• In grayscale the grayscale morphology synthesizes all pixels with a uniform intensity value.
• Conversion of training images into gray channels.
• OD-centered images in grayscale are applied to the ResNet-50 model for training.

Data Augmentation Techniques

• Data augmentation increases the number of images for statistical and biological significance.
• It is a better approach because of the limited availability of images in the medical field.
• The technique slightly modifies existing data to create more copies.
• Data augmentation also overcomes model overfitting by enhancing performance and diagnostic capability.

Transfer Learning

• Deep learning model's training requires a large image dataset, efficient hardware and more training time..
• Transfer learning uses the pre-trained model, trained on a large number of images such as ImageNet.
• Knowledge is transferred from the model to another, even if the field differs.
• The pre-trained CNN architecture ResNet-50 is retrained on the G1020, ORIGA, RIM-ONE, and DRISHTI-GS datasets.

Convolutional Neural Network

• A multilayer DL network obtaining input as high-dimension data (images) and extracts high-dimension features from input images.
• CNN architectures consist of different numbers of layers that increase as the size of the input images increases.
• Deeper networks learn more accurately, but the increase in computation time is the major drawback.
• CNNs show promising results in image processing, object detection, image segmentation, image classification, video processing, and natural language processing.

ResNet-50 Architecture

• ResNet is the short form of the residual network, and it solves the vanishing gradient problem by using the skip connection approach
• The skip connection technique in the ResNet architecture shows higher detection accuracy, takes less training time, and is easier to optimize.

Experiment Results

• The proposed model is evaluated using performance metrics: accuracy, sensitivity, and specificity.
• The four possibilities for the classified images: true positive, true negative, false positive, and false negative.
• Accuracy is the measure of the correctly labeled images divided by the total number of images.
• Sensitivity represents the correctly classified images affected by glaucoma.
• Specificity represents correctly classified healthy images.

Dataset Division and Training Details

• The dataset's fundus images were divided into three subcategories: training, validation, and testing (80%, 10%, 10%).
• Images were resized to be the same size and centered.
• The model was trained using the SDG solver with a learning rate of 0.001 on ten epochs in Python with a specified system configuration
• Four datasets were used: G1020, RIM-ONE, ORIGA, and DRISHTI-GS.
• The ResNet-50 achieved robust results with 98.48% accuracy, 96.52% specificity, 99.30% sensitivity, 97% AUC, and an F1-score of 98% on G1020 dataset.

Discussion

• The study used the deep learning architecture ResNet-50 to identify early-stage glaucoma using fundus images.
• Datasets G1020, DRISHTI-GS, RIM-ONE, and ORIGA were used for the proposed model's training, validation, and testing.
• The capability of deep learning models can automatically identify patterns from images to obtain robust results for disease detection.
• Training a model with a greater number of layers requires more time, which is not a drawback in clinical settings, as the model best classifies the images in a short amount of time.
• The use of existing models makes it possible to develop a reliable diagnosis system, despite the limited availability of the required data.

Model Performance

• The proposed model exhibited glaucoma detection on the G1020 dataset with 98.48% accuracy, 99.30% sensitivity, 96.52% specificity, an AUC of 97%, and an F1-score of 98%.
• The dataset's restricted pictures induce model overfitting, but the data augmentation technique overcomes this challenge.
• The model's performance is best on the G1020 dataset because it has a large amount of high-resolution images, but less performant on the ORIGA dataset.

Conclusions

• Glaucoma is potentially blinding and several methods have been developed for diagnosis
• The current model uses four different datasets and shows efficacy for diagnosing glaucoma at an early stage using the gray channel of fundus images.
• Self-interpretation of CNN architectures may aid clinicians in timely diagnosis and treatment of glaucoma.
• New models based on both the fundus and OCT images can be developed using a multimodal imaging approach in the future