Enhancing Eye Diseases Diagnosis through Transfer Learning: Study of Deep Convolutional Neural Networks for Accurate Classification of Glaucoma and Diabetic Retinopathy from Healthy Eye Using Fundus Images

doi:10.21203/rs.3.rs-3120228/v1

Download PDF

Article

Enhancing Eye Diseases Diagnosis through Transfer Learning: Study of Deep Convolutional Neural Networks for Accurate Classification of Glaucoma and Diabetic Retinopathy from Healthy Eye Using Fundus Images

https://doi.org/10.21203/rs.3.rs-3120228/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Fundoscopy, or ophthalmoscopy, is a medical procedure used to examine the inner structures of the eye. Fundoscopic images are valuable resources for developing artificial intelligence systems to aid in the diagnosis and management of eye conditions. This paper focuses on enhancing the robustness and generalizability of machine learning-based retinal image classification systems. A diverse and large-scale dataset of approximately 100,000 retinal images was utilized, along with a complex machine learning model. The study employed multiple datasets, including Kim's Eye Hospital, Drishti-GS1, DR HAGIS, APTOS 2019 Blindness Detection, ACRIMA, and Diabetic Retinopathy Detection 2015, to evaluate the performance of the model. Preprocessing techniques, including contrast enhancement and image resizing, were applied to prepare the dataset. The DenseNet121 model, which addresses the vanishing gradient problem, was used for transfer learning. The model created in this study can classify fundus images to three classes of Diabetic retinopathy, Glaucoma, and healthy eye with an average accuracy of 84.78%, a precision of 84.75%, and a recall of 84.76%. Although by training a model on a mild DR omitted dataset, these metrics increased significantly to an accuracy of 97.97%, a precision of 97.97%, and a recall of 97.96%. Results demonstrated that excluding mild diabetic retinopathy cases from the dataset significantly improved the model's performance.

transfer learning

deep learning

convolutional neural network

fundoscopy

diabetic retinopathy

glaucoma

Fundoscopy, also referred to as ophthalmoscopy, is a non-invasive medical procedure used to examine the innermost structures of the eye, including the retina, retinal blood vessels, optic nerve head (disk), and subjacent choroid to a limited extent. This examination provides the only means for direct inspection of the arteries, veins, and central nervous system in the patient. Fundoscopy is commonly performed by medical professionals such as ophthalmologists, internists, neurologists, or pediatricians to diagnose and monitor various eye conditions like macular degeneration, diabetic retinopathy, retinal hemorrhages, papilledema, and glaucoma [1].

During the fundoscopic examination, short-acting topical parasympatholytic eye drops are frequently used to dilate the patient's pupils. A special optic device, called an ophthalmoscope, is then used to illuminate the retina through the pupil. The light rays then form an image of the retina which is visible through the pupil. The results of the fundoscopic examination can provide valuable information about the overall health of the eye and the presence of various abnormalities or diseases [1].

Fundoscopic images can also be taken during the examination to document any abnormalities or changes in the eye. To take fundoscopic images, a camera is attached to the ophthalmoscope, and the images are captured by focusing the camera on the retina or other structures of the eye [2]. Collecting fundoscopic images is a valuable resource for developing artificial intelligence (AI) systems that can aid in the diagnosis and management of various eye conditions. Multiple machine learning (ML) researchers have already utilized well-known public ophthalmological imaging datasets, such as MESSIDOR, DRIVE, EyePACS, and E-ophtha, on several occasions [3–6].

By feeding large datasets of fundoscopic images into AI algorithms, computer systems can learn to identify and analyze patterns in the images that correspond to specific eye conditions. These AI systems have the potential to aid medical professionals in diagnosing and treating eye diseases more effectively, by enhancing accuracy and efficiency, while saving time and providing better access to care. Additionally, these systems can standardize treatment procedures, develop personalized treatment plans, and improve overall patient outcomes.

Diabetic retinopathy (DR) is a common complication of diabetes mellitus that affects the eyes. It is the leading cause of vision loss in middle-aged and elderly people and is a significant public health concern worldwide [7]. Among adults with diabetes in the US, the prevalence of diabetic retinopathy and vision-threatening diabetic retinopathy (VTDR) is 28.5% and 4.4%, respectively [8]. It is projected that by 2045, the prevalence of DR and VTDR worldwide will increase to approximately 160.50 million and 44.82 million, respectively. This rise will have a greater impact on people with diabetes residing in the Middle East and North Africa, as well as the Western Pacific regions [9].

The identification of certain features in the retina, such as microaneurysms, hemorrhages, and exudates, is used to diagnose DR. Clinically, it is classified as mild, moderate and severe non-proliferative DR and proliferative DR based on the extent of retinopathy [7]. Screening and timely treatment of retinal sight threatening complications have been shown to reduce blindness due to these complications in the countries where these services are well established such as the United Kingdom [10]. Retinal photography accompanied by manual interpretation is a commonly utilized screening method for diabetic retinopathy. This technique has been shown to perform better than in-person dilated eye exams in some cases [11].

Most low- and middle-income countries lack well-established primary care infrastructures, making it difficult to carry out systematic screening programs for diabetic retinopathy. As a result, there is an urgent need for tailored, affordable screening and treatment plans to serve a significant portion of the population living with diabetes [10]. Deep learning (DL) is a powerful machine learning technique that can automatically learn the most significant features directly from images, given a substantial dataset of labeled examples [5]. DL has shown impressive results of high sensitivity and specificity in automated image analysis of fundus photographs. Utilizing deep neural networks provides a significant advantage in screening for diabetic retinopathy through retinal images. This approach can enhance the identification of diabetic retinopathy lesions and risk factors with a high level of accuracy and reliability [10].

Glaucoma is another eye condition that, if left untreated or inadequately managed, can lead to vision Loss and blindness. It is the leading cause of irreversible blindness worldwide and is associated with a reduced quality of life [12]. Glaucoma is a group of progressive optic neuropathies characterized by a degeneration of retinal ganglion cells and retinal nerve fiber layers that result in changes in the optical nerve head [13]. Glaucoma is associated with intraocular pressure (IOP)-related damage to the optic nerve, which results in the loss of retinal ganglion cells [14].

The prevalence of glaucoma tends to increase with age, and it is more common in older adults. According to the World Health Organization (WHO), it is estimated that over 76 million people were affected by glaucoma in 2020 and it is estimated that the number of individuals aged 40–80 years with glaucoma will reach 111.8 million by 2040 [15]. The global prevalence of primary open-angle glaucoma (POAG), the most common type of glaucoma, is estimated to be 2.2%, impacting approximately 57.5 million people. In Europe, the prevalence of POAG is 2.51%, affecting around 7.8 million individuals. In the UK, POAG affects 2% of people older than 40 years and 10% of those older than 75 years, with a higher prevalence observed in African-Caribbean individuals. Primary angle-closure glaucoma (PACG) is less prevalent, affecting 0.17% of individuals younger than 40 years, particularly East Asians [16].

Currently, diagnosing and monitoring glaucoma necessitate a comprehensive eye examination and the collection of various data, which can be complex to interpret. Additionally, there is considerable similarity in ocular features between individuals without glaucoma and those with early-stage glaucoma [17, 18]. Consequently, there is a growing interest in developing complementary techniques, such as artificial intelligence (AI) systems, to aid in diagnosis and screening progression of glaucoma.

Key tools in the diagnosis of glaucoma include Heidelberg retinal tomography (HRT), optical coherence tomography (OCT), visual field (VF) tests, and color fundus photography with red-free imaging [19]. A web-based telemedicine program named “DYSEO” was developed to screen for Glaucoma, focusing on evaluating the effectiveness of OCT and retinography in detecting the disease. The study successfully identified approximately 20% of individuals as potential glaucoma suspects among those initially examined. The detection rate for confirmed glaucoma cases was 4.9%, while it reached 9.2% for confirmed or probable glaucoma. Overall, the diagnostic accuracy of the screening program was found to be moderate to high [20]. By incorporating ML and AI techniques into the telemedicine screening programs, the diagnostic accuracy can be further improved, detection rates can increase, and the overall efficiency of the program can be enhanced, leading to better identification and management of glaucoma cases.

Many researchers in the field of glaucoma detection have faced limitations due to working with small datasets and fundus images from private sources, which often lack real-time variations in image quality. This situation hinders the robustness and generalizability of the developed models [21, 22]. To address this challenge, it is crucial to develop a model that can effectively handle images acquired under different environmental conditions.

This article focuses on enhancing the robustness and generalizability of machine learning-based retinal image classification systems. To achieve this, we utilized a Diverse and Large-Scale Dataset consisting of approximately 100,000 retinal images. Furthermore, we employed a very complex machine learning model capable of effectively handling such a vast amount of data. By combining this extensive dataset and powerful model, we aimed to significantly improve the performance and reliability of retinal image classification systems.

Datasets

Kim's Eye Hospital

A collection of fundus photographs from both normal and glaucoma patients was gathered from Kim's Eye Hospital. In total, 1,544 photos were acquired, with 788 photos from normal patients and 756 photos from glaucoma patients. These images were standardized to a size of 240x240 pixels and focused solely on the optic disc region of the retina [23].

Drishti-GS1 dataset

The Drishti-GS1 dataset, provided by the Medical Image Processing (MIP) group at IIIT Hyderabad, comprises a total of 101 images. Out of these, 31 images were obtained from normal patients, while 70 images were from patients diagnosed with glaucoma. Each image in the dataset has been annotated by four eye experts with varying clinical experience. All the images were collected at Aravind Eye Hospital in Madurai, with the informed consent of the hospital visitors. The selection of glaucoma patients was conducted by clinical investigators based on clinical findings during the hospital visit. The selected patients ranged in age from 40 to 80 years, with an approximately equal distribution of males and females [24].

DR HAGIS database

The DR HAGIS database was specifically designed to assist in the development of vessel extraction algorithms applicable to retinal screening programs. 20 fundus images were obtained from this database, comprising 10 images of glaucoma and 10 images of diabetic retinopathy [25].

APTOS 2019 Blindness Detection

We utilized the APTOS 2019 dataset, which was obtained from the Kaggle competition and provided by Aravind Eye Hospital in India. The primary goal of the research was to detect and prevent diabetic retinopathy in rural areas where conducting medical screenings is challenging. To achieve this, technicians from Aravind Eye Hospital visited these rural areas and captured images using specialized fundus cameras equipped with high-resolution capabilities and attached microscopes. Subsequently, highly skilled doctors reviewed the captured images to provide accurate diagnoses. The dataset consisted of 3,662 photos in total, with 1,805 photos from normal patients and 1,857 photos from DR patients [26].

ACRIMA Database

The ACRIMA database comprises a total of 705 fundus images, with 396 images representing glaucomatous cases and 309 images depicting normal cases. These images were collected at the FISABIO Oftalmología Médica in Valencia, Spain, from patients who had provided their prior informed consent. The collection process followed the ethical standards outlined in the 1964 Declaration of Helsinki. To ensure accuracy, all images in the ACRIMA database were annotated by glaucoma experts with extensive years of experience. Additionally, the images were cropped around the optic disc and assigned new names for identification purposes [27].

Diabetic Retinopathy Detection 2015

The dataset consists of images captured from different models and types of cameras, which can result in variations in the visual appearance of left and right retinas. Some images display the retina anatomically, with the macula on the left and the optic nerve on the right for the right eye. Other images simulate the view through a microscope condensing lens, appearing inverted compared to a typical live eye exam. There are two main indicators to determine if an image is inverted: the position of the macula relative to the midline of the optic nerve, and the presence or absence of a notch on the side of the image. However, it is important to note that the dataset may contain noise, such as artifacts, focus issues, underexposure, or overexposure. Retinal images were provided by EyePACS, a free platform for retinopathy screening [28]. The dataset was obtained from the Kaggle competition [29]. The objective of this competition was to develop robust algorithms capable of functioning effectively in the presence of such noise and variations. The dataset consisted of 88,702 photos in total, with 65,343 photos from normal patients and 23,359 photos from DR patients (Table 1).

Table 1
	Healthy eye	Glaucoma	Diabetic retinopathy	Total
Kim's Eye Hospital	788	756	0	1,544
Drishti-GS1	31	70	0	101
DR HAGIS	0	10	10	20
APTOS 2019	1,805	0	1,857	3,662
ACRIMA	309	396	0	705
DR Detection 2015	65,343	0	23,359	88,702
Total	68,276	1,232	25,226	94,734
Description of all datasets used in this study.

Study design and workflow

Image Preprocessing:

In order to label the collected dataset, we utilized the metadata information associated with each image. The dataset consisted of three categories: Healthy eye, Diabetic retinopathy, and Glaucoma. The labeling process involved extracting the image category information from the metadata file and subsequently organizing the images into folders corresponding to their respective categories.

To enhance the image details and extract valuable information, we employed the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. Our approach entailed converting each image from RGB to LAB color space, applying CLAHE specifically on the L (lightness) channel using a window size of 8x8, and then converting the image back to the RGB color mode. This technique effectively enhanced the contrast of the images and made the details more prominent, aiding in the subsequent analysis [30].

Since we opted to use the DenseNet121 model for classification, certain preprocessing steps were necessary to prepare the dataset for input to the model. Firstly, we converted the images from the RGB color mode to the BGR color mode, as per the requirements of the model. Subsequently, we resized all the images to a uniform size of 224x224 pixels to ensure consistency. Finally, we performed zero-centering on the images, which is a common preprocessing technique in model building [31].

These preprocessing techniques were employed to ensure that the dataset was appropriately prepared and aligned with the requirements of the DenseNet121 model. By applying these steps, we aimed to optimize the performance of the model and facilitate accurate classification of the eye images.

Model Architecture for Transfer Learning:

Convolutional Neural Networks (CNNs) have emerged as a popular approach for image classification tasks, primarily due to their ability to effectively extract both global and local features from images. In a conventional feed-forward CNN architecture, the output of each convolutional layer serves as the input to the subsequent layer, thereby progressively capturing higher-level features.

However, as the depth of the CNN model increases by adding more layers, a phenomenon known as the vanishing gradient problem can arise. The vanishing gradient problem occurs when the information flow between the input and output layers becomes extended, potentially resulting in the loss of crucial information. This issue can significantly impact the efficiency of the training process [31].

When gradients become vanishingly small during backpropagation, the model's ability to learn and update the weights diminishes, leading to slower convergence and suboptimal performance. Consequently, deeper CNN models may struggle to learn complex patterns and exhibit reduced training efficiency [32].

DenseNet architectures address the vanishing gradient problem by introducing a unique connectivity pattern among layers. Unlike conventional CNNs where each layer is connected only to the subsequent layer, DenseNet models establish dense connections between all layers (Fig. 1). This design choice facilitates the seamless flow of information throughout the network and offers several advantages in terms of accuracy and performance [33].

In addition to addressing the vanishing gradient problem, another important consideration when choosing a DenseNet model is its suitability for real-time applications and its overall computational efficiency. With these factors in mind, we have selected DenseNet121 as the pre-trained model for our classification problem. DenseNet121 is a variant of DenseNet that strikes a balance between model complexity and performance. It has been specifically designed to provide a lightweight architecture that is well-suited for real-time applications [34].

Experimental Setups, Model Development, and Performance Measurement:

We implemented the proposed classification model on a Google Colaboratory environment with an Intel(R) Xeon(R) 2.30GHz CPU processor and 13GB of RAM and NVIDIA Tesla T4 CUDA enabled GPU processor with CUDA 11.2 which has designed for high-performance computing, deep learning training and inference, machine learning, and data analytics. The model was created with Python 3.10, TensorFlow 2.11, Scikit-Learn 1.0.2, and Numpy as dependencies.

Upon loading the pre-trained DenseNet121 model, we proceeded to modify its classifier component by adding our own customized classifier. This classifier includes a global average pooling layer, a dropout layer with a dropout rate of 0.5, and a dense layer with three units that utilizes the softmax activation function. The architecture of the proposed model can be visualized in Fig. 2.

To compile the model, we employed the categorical cross-entropy loss function, which is well-suited for multiclass classification tasks, and the Adam optimizer with an initial learning rate of 5e-5. This combination helps optimize the model's performance during the training process.

In order to fine-tune the pre-trained DenseNet121 model to our specific task, we made the layers of the pre-trained model trainable. This allows the model to adapt and learn from our dataset, while also leveraging the knowledge and representations acquired from the pre-training phase. By utilizing the pre-trained weights as the starting point, we provide the model with a strong initialization, which can help accelerate the convergence and enhance overall performance. The training flow process is shown in Fig. 3.

The dataset was divided into training and testing sets, with approximately two-thirds of the data points allocated to the training set and the remaining one-third to the testing set. This partitioning allows us to train the model on a large portion of the data while reserving a separate set for evaluating its performance on unseen samples. To further enhance the robustness of our results, we applied 10-fold cross-validation. This technique involves dividing the dataset into ten equally sized subsets or "folds." The model is trained and evaluated ten times, each time using a different fold as the testing set while the remaining nine folds are used for training.

Given the imbalance in our dataset, where the number of samples in each class varies significantly, relying solely on accuracy as an evaluation metric may not provide a comprehensive assessment of the model's performance. To address this, we employed additional metrics such as recall, precision, and AUC (Area Under the ROC Curve) for each class.

In this section, we present the experimental results of our models trained on the whole combined dataset and a dataset without mild DR. We observed that errors primarily occurred in the misclassification of mild DR as normal, which could be attributed to the CNN's inability to detect subtle disease features. To assess the impact of excluding mild DR cases on the results, we also trained a model on a dataset without mild DR and analyzed its performance.

The model underwent 10-fold training using the Adaptive Momentum (Adam) optimizer with the Categorical Cross Entropy loss function. We initialized the learning rate to 5e-5 and set all layers to be trainable during training. The training process involved all layers of the model. To dynamically adjust the learning rate, we utilized the ReduceLROnPlateau callback, which reduced the learning rate if the validation loss did not improve after 3 epochs. The recorded metrics after each fold of training are summarized in Table 2 and Fig. 4.

Table 2
Fold	Whole Dataset Model				Mild DR Class Omitted Dataset Model
Fold	Loss	Accuracy	Precision	Recall	Loss	Accuracy	Precision	Recall
1	0.386	0.848	0.848	0.847	0.064	0.982	0.982	0.982
2	0.369	0.853	0.852	0.852	0.082	0.977	0.977	0.977
3	0.384	0.844	0.844	0.844	0.071	0.977	0.977	0.977
4	0.398	0.836	0.837	0.835	0.078	0.974	0.974	0.974
5	0.391	0.852	0.852	0.852	0.051	0.981	0.981	0.981
6	0.379	0.848	0.848	0.848	0.069	0.976	0.976	0.976
7	0.403	0.844	0.844	0.844	0.050	0.982	0.982	0.982
8	0.386	0.849	0.849	0.848	0.070	0.981	0.981	0.981
9	0.372	0.851	0.851	0.851	0.068	0.980	0.980	0.980
10	0.377	0.849	0.850	0.849	0.065	0.980	0.980	0.980
Mean	0.3845	0.847	0.847	0.847	0.0668	0.979	0.979	0.979
Model’s performance on the test set at the end of each fold of training.

To comprehensively evaluate the model's performance across all classes and assess its ability to discriminate between different classes, a Multiclass ROC (Receiver Operating Characteristic) analysis was conducted using the one-vs-all method. This technique extends the binary classification approach to multiclass problems.

In the analysis, the true positive rate (TPR) and false positive rate (FPR) were calculated for each class individually. The model's predictions for that specific class were considered as positive, while the predictions for all other classes were treated as negative. This process was repeated for each class, resulting in a separate ROC curve generated for each class at the end of each fold (Fig. 5).

As depicted in Fig. 5, the removal of the mild DR images from the dataset led to a significant increase in the AUC of the ROC curve. The macro average AUC for the model trained on the complete dataset is 0.905, whereas the model trained on the dataset with mild DR omitted images achieved an AUC of 0.990.

Additionally, when comparing the performance metrics of the two models, it is evident that the model trained on the dataset without mild DR images exhibited notable improvements. The macro average accuracy, precision, and recall of the model trained on the complete dataset were 84.74%, 84.75%, and 84.76% respectively. However, in the model trained on the mild DR omitted dataset, these metrics increased significantly to 97.97%, 97.97%, and 97.96% respectively. These results demonstrate our model’s weakness in differentiating between mild diabetic retinopathy and healthy eye.

To identify and analyze common types of errors made by the model, such as false positives and false negatives, confusion matrices were plotted, as shown in Fig. 6. The confusion matrix provides a tabular representation of the model's predictions compared to the actual true labels of the dataset. It displays the proportion of true positive, true negative, false positive, and false negative predictions made by the model for each class.

In our analysis, we observed that after removing mild DR images from the dataset, the model's performance improved not only in predicting DR but also in predicting glaucoma and healthy eyes. This improvement can be seen in the increased true positive rates across all classes. These findings suggest that the presence of mild DR images in the dataset can also interfere with the accurate prediction of glaucoma. By removing these images, the model's ability to correctly classify glaucoma cases was enhanced.

AI has brought about a revolutionary transformation in research, optimizing and streamlining processes. The abundance of patient information stored in electronic medical records (EMRs) presents a challenge in converting this data into actionable knowledge that can benefit clinicians, researchers, and healthcare systems in providing precise, personalized, and efficient care. Fortunately, AI has emerged as a solution, enabling the development of computer-generated algorithms and automated systems that expedite data collection [35].

The field of ophthalmology, with its access to ophthalmic imaging and objective markers, is particularly well-suited for leveraging large datasets. These datasets not only facilitate the identification of correlations but also enable multicenter studies, support multimodal analyses, unveil novel imaging patterns, and enhance the statistical power of research. Often referred to as "big data," this extensive collection of medical information serves as an ideal foundation for the application of AI, machine learning (ML), and deep learning (DL) algorithms on an unprecedented scale [36].

In this study, we introduce an extraction method that utilizes transfer learning to extract fundus image features from the DenseNet-121 model. The extracted features are subsequently fed into a stack of dense layers, with a SoftMax activation function at the final layer, enabling multiclass classification of retinal images into three distinct classes: Healthy Eye, Glaucoma, and Diabetic Retinopathy.

Novel Contributions

Extensive research has already been conducted, with over 30 papers focusing on Diabetic retinopathy diagnosis and over 20 papers on glaucoma diagnosis using machine learning approaches for diagnosis. These studies utilized different datasets with varying sample sizes and employed diverse methodologies. To further advance in the field, we embarked on a unique approach. Instead of relying on individual datasets, we made the decision to combine multiple datasets, allowing us to create the largest dataset possible. By doing so, we could train a single model capable of simultaneously diagnosing both glaucoma and Diabetic retinopathy from healthy eye images. This approach represents a significant departure from previous studies, which typically focused on one condition at a time.

The integration of multiple datasets not only expands the scope of our research but also enables us to leverage a broader range of image variations, improving the robustness and generalizability of our model. Through this research, we aim to enhance the accuracy and efficiency of diagnosis for both glaucoma and Diabetic retinopathy, paving the way for more comprehensive and effective clinical applications.

State of the art

In this section, we present an overview of current advancements, techniques, methodologies, and research findings relevant to our study. This review provides a comprehensive snapshot of the state of the art in this research area.

In 2018, Lam et al. developed Convolutional Neural Network (CNN) models for the purpose of recognizing and staging diabetic retinopathy. Their study involved the utilization of two fundoscopic image datasets: the Kaggle Diabetic Retinopathy Images dataset and the Messidor-1 dataset [37].

In order to enhance the learning of challenging features, the researchers transitioned to a smaller but more suitable dataset. To achieve this, they supplemented the Messidor dataset with a Kaggle partition called "MildDR," which comprised of 550 images. By combining these datasets, they created a new dataset consisting of 1200 retinal images classified into four distinct classes [37].

They discovered that preprocessing with contrast limited adaptive histogram equalization on 3 class classifier increases sensitivity from 0 to 29.4% in the mild class while this measure was approximately the same for the remaining two classes. They used AlexNet, VGG16 and GoogLeNet models for the Binary model and The GoogLeNet model achieved the highest sensitivity of 95% and specificity of 96%. However, when they trained 3-ary and 4-ary classifiers with a GoogLeNet model on the Kaggle dataset, they were unable to achieve significant sensitivity levels for the mild class. Same as their research we also encountered that errors primarily occur in the misclassification of mild disease as normal due to the CNNs inability to detect subtle disease features [37].

During their research, the authors made an intriguing discovery regarding the preprocessing technique known as contrast limited adaptive histogram equalization (CLAHE) in the context of a three-class classifier. They found that applying CLAHE increased the sensitivity in the mild class from 0–29.4%, while the impact on the remaining two classes was negligible [37].

For their binary model, the researchers employed three popular CNN architectures: AlexNet, VGG16, and GoogLeNet. Among these models, GoogLeNet demonstrated the highest sensitivity of 95% and specificity of 96%. However, when training 3-ary and 4-ary classifiers using the GoogLeNet model on the Kaggle dataset, the researchers were unable to achieve significant sensitivity levels for the mild class. Interestingly, their findings aligned with our own research, as we also encountered challenges related to the misclassification of mild disease as normal. This was primarily attributed to CNN’s limited ability to detect subtle disease features [37].

In a study conducted by Li et al. in 2019, the GoogLeNet model was utilized for the task of diabetic retinopathy grading. They collected a dataset consisting of 13,676 retinal images and evaluated the performance of various models. Interestingly, the ResNet-18 model exhibited superior performance compared to other models, achieving an average accuracy of 76.59% on the test set. This surpassed the average accuracy of 74.98% achieved by the GoogLeNet model [38].

ResNet, short for Residual Neural Network, is a deep learning architecture that was introduced by Kaiming He et al. in 2015 and won the ImageNet 2015 competition. Its primary goal was to address the challenge of training very deep neural networks, which had previously been difficult due to the problem of vanishing gradients. The key innovation of ResNet lies in its use of residual blocks, which employ shortcut connections or skip connections within the network [39].

ResNets has been widely utilized for the classification task of DR. Zhang et al. (2019) conducted a study using ResNet on 3,833 macula-centered retinal fundus images. However, the original dataset contained a larger number of images, comprising 13,767 images from 1,872 patients, which included fluorescence contrast images and postoperative fundus images. The authors excluded these additional images, resulting in 9,934 images from 1,229 patients to be excluded [40].

In their study, Zhang et al. compared the performance of various deep learning models, including ResNet50, DenseNets, InceptionV3, InceptionV2, and Xception, for DR classification. They reported classification accuracies of 94.94%, 93.34%, 94.55%, 94.16%, and 92.09%, respectively, for these models. Among them, Xception exhibited the lowest classification accuracy, while ResNet achieved the highest accuracy. Interestingly, in the identification task, Xception achieved the highest accuracy, while ResNet had the lowest accuracy in this study [40].

Gulshan et al. conducted a study where they employed a pre-trained InceptionV3 model to classify diabetic retinopathy (DR) using a dataset of 128,175 images. The images were retrospectively collected from EyePACS in the US and three eye hospitals in India. The study also included two clinical validation sets: EyePACS-1 and Messidor-2, containing 9,963 and 1,748 images, respectively [41].

For the high specificity operating point, the algorithm achieved a sensitivity of 90.3% and 87.0% and a specificity of 98.1% and 98.5% in EyePACS-1 and Messidor-2, respectively. The referable diabetic retinopathy was defined as moderate or worse diabetic retinopathy or referable macular edema, as determined by a panel of at least seven US board-certified ophthalmologists. At the high sensitivity operating point, the algorithm demonstrated a sensitivity of 97.5% and 96.1% and a specificity of 93.4% and 93.9% in the two validation sets, respectively [41].

The development of new technologies has increasingly shown their relevance in the field of glaucoma diagnosis and treatment. In particular, ML techniques have emerged as essential tools for achieving significant research outcomes. Creating a CNN model for glaucoma detection using retina fundus images offers numerous advantages, including early detection, objectivity, scalability, and enhanced diagnostic accuracy [42].

Li et al. conducted a study in which they developed a deep learning system specifically designed for classifying glaucomatous optic neuropathy using color fundus photographs. The research team recruited 21 trained ophthalmologists who could classify the photographs. A total of 48,116 fundus images were included in the study, out of which 39,745 images met the criteria for training and validation [43].

The developed deep convolutional neural network consisted of 22 layers. When evaluated on the validation dataset, the system achieved an impressive accuracy of 92.9% in classifying referable glaucomatous optic neuropathy. Additionally, it demonstrated a sensitivity of 95.6% and a specificity of 92.0%, indicating its ability to effectively identify cases of glaucomatous optic neuropathy [43].

Fu et al. proposed a novel approach called Disc-aware Ensemble Network (DENet) for automated glaucoma screening. The DENet model was designed to consider both global and local image levels to enhance its screening performance [44].

At the global level, the DENet model consisted of two streams. The first stream utilized a standard classification network based on ResNet. The second stream employed a segmentation-guided network adapted from the U-shape convolutional network. This segmentation-guided network helped in identifying and localizing important regions within the image. The DENet model achieved promising results with an accuracy of 84.29%, a specificity of 83.80%, and a sensitivity of 84.78% on SCES Dataset [44].

Raghavendra et al. conducted a study to develop a glaucoma detection method utilizing an 18-layer CNN architecture. The architecture included convolutional and max-pool layers, followed by a classification layer employing the logarithmic soft-max activation function. The method was evaluated using a dataset of 1,426 fundus images, comprising 589 normal images and 837 glaucoma images. Remarkably, the proposed method achieved the highest accuracy of 98.13% in classifying the fundus images into normal and glaucoma categories [45].

In 2018, Christopher et al. conducted a study in which they recruited participants from two longitudinal studies focused on assessing optic nerve structure and visual function in glaucoma: the African Descent and Glaucoma Evaluation Study (ADAGES) and the University of California, San Diego (UCSD) Diagnostic Innovations in Glaucoma Study (DIGS). The datasets comprised stereo fundus images obtained from these studies. The dataset included 14,822 individual optic nerve head (ONH) images from 4,363 eyes of 2,329 individuals [46].

Three deep learning architectures, VGG16, Inception, and ResNet50, were evaluated using both native and transfer learning approaches. The models were trained, validated, and tested using a 10-fold cross-validation setup. The best performing model was determined based on the area under the receiver operating characteristic curve (AUC) and was found to be the transfer learning ResNet50 model, achieving an AUC of 0.91 [46].

The model created in this study is able to classify fundus images to three classes of Diabetic retinopathy, Glaucoma, and healthy eye with an average accuracy of 84.78%, a precision of 84.75%, and a recall of 84.76%. Although by training a model on a mild DR omitted dataset, these metrics increased significantly to an accuracy of 97.97%, a precision of 97.97%, and a recall of 97.96%.

Weaknesses

The class distribution in our combined dataset was imbalanced, with varying numbers of images in each class. This can impact the performance of the classification model, particularly for the minority class glaucoma. While our model shows promising results in terms of classification accuracy, it is essential to validate its performance in real clinical settings. Clinical validation studies involving independent datasets and assessments by ophthalmologists are necessary to assess the model's generalizability. Although we employed cross-validation techniques to assess the performance of our model on different subsets of the combined dataset, our model would benefit from external validation on completely independent datasets.

This study utilized transfer learning and a combined dataset to develop a classification model for diagnosing glaucoma and diabetic retinopathy from fundus images. Multiple datasets, including Kim's Eye Hospital, Drishti-GS1, DR HAGIS, APTOS 2019 Blindness Detection, ACRIMA, and Diabetic Retinopathy Detection 2015(EyePACS), were combined to create a comprehensive and diverse dataset for training the model.We employed the DenseNet121 architecture, which addresses the vanishing gradient problem and offers computational efficiency suitable for real-time applications. Preprocessing techniques such as contrast enhancement, scaling, and image resizing were applied to optimize the dataset for input to the model.

Results demonstrated that excluding mild diabetic retinopathy (DR) cases from the dataset significantly improved the model's performance, particularly in terms of the area under the ROC curve (AUC). The model trained on the complete dataset achieved a macro average AUC of 0.905, while the model trained on the dataset without mild DR reached an AUC of 0.990. Moreover, the model trained on the mild DR omitted dataset exhibited substantial improvements in accuracy, precision, and recall across all classes compared to the model trained on the complete dataset.

Ethics concerning human participants

As we did not directly deal with human subjects and used anonymized and unidentifiable data from previously published and openly accessible studies, we declare that the ethical concerns do not apply to this study.

Data availability

The datasets used during the current study are publicly available and can be downloaded from the following links:

Kim's Eye Hospital dataset: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/1YRRAC
Drishti-GS1 dataset: https://www.kaggle.com/datasets/lokeshsaipureddi/drishtigs-retina-dataset-for-onh-segmentation
DR HAGIS database: personalpages.manchester.ac.uk/staff/niall.p.mcloughlin/
APTOS 2019 Blindness Detection dataset: https://www.kaggle.com/competitions/aptos2019-blindness-detection/data
ACRIMA Database: https://figshare.com/s/c2d31f850af14c5b5232
Diabetic Retinopathy Detection 2015 dataset: https://www.kaggle.com/competitions/diabetic-retinopathy-detection/data

Author contributions

A.T., M.K., and M.T built and trained models. A.T., and A.A.H performed the statistical analyses. A.T. designed figures and tables. A.T. and M.K. directed the study. All authors reviewed the manuscript.

Competing interests

The authors declare no competing interests.

Walker HK, Hall WD, Hurst JW. Clinical methods: the history, physical, and laboratory examinations. 1990.
Mishra C, Tripathy K. Fundus Camera. StatPearls [Internet]: StatPearls Publishing; 2022.
Decencière E, Zhang X, Cazuguel G, Laÿ B, Cochener B, Trone C, et al. FEEDBACK ON A PUBLICLY DISTRIBUTED IMAGE DATABASE: THE MESSIDOR DATABASE. Image Analysis & Stereology. 2014:231-4.
Naveed K, Abdullah F, Madni HA, Khan MAU, Khan TM, Naqvi SS. Towards Automated Eye Diagnosis: An Improved Retinal Vessel Segmentation Framework Using Ensemble Block Matching 3D Filter. Diagnostics (Basel). 2021;11(1).
Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. Jama. 2016;316(22):2402-10.
Decencière E, Cazuguel G, Zhang X, Thibault G, Klein JC, Meyer F, et al. TeleOphta: Machine learning and image processing methods for teleophthalmology. IRBM. 2013;34(2):196-203.
Wong TY, Cheung CM, Larsen M, Sharma S, Simó R. Diabetic retinopathy. Nat Rev Dis Primers. 2016;2:16012.
Zhang X, Saaddine JB, Chou CF, Cotch MF, Cheng YJ, Geiss LS, et al. Prevalence of diabetic retinopathy in the United States, 2005-2008. Jama. 2010;304(6):649-56.
Teo ZL, Tham YC, Yu M, Chee ML, Rim TH, Cheung N, et al. Global Prevalence of Diabetic Retinopathy and Projection of Burden through 2045: Systematic Review and Meta-analysis. Ophthalmology. 2021;128(11):1580-91.
Raman R, Srinivasan S, Virmani S, Sivaprasad S, Rao C, Rajalakshmi R. Fundus photograph-based deep learning algorithms in detecting diabetic retinopathy. Eye (Lond). 2019;33(1):97-109.
Chakrabarti R, Harper CA, Keeffe JE. Diabetic retinopathy management guidelines. Expert Review of Ophthalmology. 2012;7(5):417-39.
Thomas S, Hodge W, Malvankar-Mehta M. The Cost-Effectiveness Analysis of Teleglaucoma Screening Device. PLoS One. 2015;10(9):e0137913.
Harasymowycz P, Birt C, Gooi P, Heckler L, Hutnik C, Jinapriya D, et al. Medical Management of Glaucoma in the 21st Century from a Canadian Perspective. J Ophthalmol. 2016;2016:6509809.
Weinreb RN, Leung CK, Crowston JG, Medeiros FA, Friedman DS, Wiggs JL, et al. Primary open-angle glaucoma. Nat Rev Dis Primers. 2016;2:16067.
Tham YC, Li X, Wong TY, Quigley HA, Aung T, Cheng CY. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology. 2014;121(11):2081-90.
Allison K, Patel D, Alabi O. Epidemiology of Glaucoma: The Past, Present, and Predictions for the Future. Cureus. 2020;12(11):e11686.
Akter N, Fletcher J, Perry S, Simunovic MP, Briggs N, Roy M. Glaucoma diagnosis using multi-feature analysis and a deep learning technique. Scientific Reports. 2022;12(1):8064.
Greenfield DS, Weinreb RN. Role of optic nerve imaging in glaucoma clinical practice and clinical trials. Am J Ophthalmol. 2008;145(4):598-603.
Michelessi M, Lucenteforte E, Oddone F, Brazzelli M, Parravano M, Franchi S, et al. Optic nerve head and fibre layer imaging for diagnosing glaucoma. Cochrane Database Syst Rev. 2015;2015(11):Cd008803.
Anton A, Nolivos K, Pazos M, Fatti G, Ayala ME, Martínez-Prats E, et al. Diagnostic Accuracy and Detection Rate of Glaucoma Screening with Optic Disk Photos, Optical Coherence Tomography Images, and Telemedicine. J Clin Med. 2021;11(1).
Kanse SS, Yadav DM. Retinal Fundus Image for Glaucoma Detection: A Review and Study. 2019;28(1):43-56.
Shinde R. Glaucoma detection in retinal fundus images using U-Net and supervised machine learning algorithms. Intelligence-Based Medicine. 2021;5:100038.
Kim U. processed_data.zip. Machine learn for glaucoma. V1 ed: Harvard Dataverse; 2018.
Sivaswamy J, Krishnadas SR, Joshi GD, Jain M, Tabish AUS, editors. Drishti-GS: Retinal image dataset for optic nerve head(ONH) segmentation. 2014 IEEE 11th International Symposium on Biomedical Imaging (ISBI); 2014 29 April-2 May 2014.
Holm S, Russell G, Nourrit V, McLoughlin N. DR HAGIS-a fundus image database for the automatic extraction of retinal surface vessels from diabetic patients. J Med Imaging (Bellingham). 2017;4(1):014503.
Karthick M, Sohier D. APTOS 2019 Blindness Detection. Kaggle https://kaggle com/competitions/aptos2019-blindness-detection Go to reference in chapter. 2019.
Diaz-Pinto A, Morales S, Naranjo V, Köhler T, Mossi JM, Navea A. CNNs for automatic glaucoma assessment using fundus images: an extensive validation. Biomed Eng Online. 2019;18(1):29.
Bhaskaranand M, Cuadros J, Ramachandra C, Bhat S, Nittala MG, Sadda SR, et al., editors. EyeArt+ EyePACS: automated retinal image analysis for diabetic retinopathy screening in a telemedicine system. Ophthalmic Medical Image Analysis International Workshop; 2015: University of Iowa.
Graham B. Kaggle diabetic retinopathy detection competition report. University of Warwick. 2015:24-6.
Pisano ED, Zong S, Hemminger BM, DeLuca M, Johnston RE, Muller K, et al. Contrast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms. J Digit Imaging. 1998;11(4):193-200.
Hasan N, Bao Y, Shawon A, Huang Y. DenseNet convolutional neural networks application for predicting COVID-19 using CT image. SN computer science. 2021;2(5):389.
Taye MM. Theoretical Understanding of Convolutional Neural Network: Concepts, Architectures, Applications, Future Directions. Computation [Internet]. 2023; 11(3).
Huang G, Liu Z, Van Der Maaten L, Weinberger KQ, editors. Densely connected convolutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition; 2017.
Chauhan T, Palivela H, Tiwari S. Optimization and fine-tuning of DenseNet model for classification of COVID-19 cases in medical imaging. 2021;1(2):100020.
Srivastava O, Tennant M, Grewal P, Rubin U, Seamone M. Artificial intelligence and machine learning in ophthalmology: A review. Indian J Ophthalmol. 2023;71(1):11-7.
Chiang MF, Sommer A, Rich WL, Lum F, Parke DW, 2nd. The 2016 American Academy of Ophthalmology IRIS(®) Registry (Intelligent Research in Sight) Database: Characteristics and Methods. Ophthalmology. 2018;125(8):1143-8.
Lam C, Yi D, Guo M, Lindsey T. Automated Detection of Diabetic Retinopathy using Deep Learning. AMIA Jt Summits Transl Sci Proc. 2018;2017:147-55.
Li T, Gao Y, Wang K, Guo S, Liu H, Kang H. Diagnostic assessment of deep learning algorithms for diabetic retinopathy screening. Information Sciences. 2019;501:511-22.
He K, Zhang X, Ren S, Sun J, editors. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition; 2016.
Zhang W, Zhong J, Yang S, Gao Z, Hu J, Chen Y, et al. Automated identification and grading system of diabetic retinopathy using deep neural networks. Knowledge-Based Systems. 2019;175:12-25.
Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. Jama. 2016;316(22):2402-10.
Cabitza F, Rasoini R, Gensini GF. Unintended Consequences of Machine Learning in Medicine. Jama. 2017;318(6):517-8.
Li Z, He Y, Keel S, Meng W, Chang RT, He M. Efficacy of a Deep Learning System for Detecting Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. Ophthalmology. 2018;125(8):1199-206.
Fu H, Cheng J, Xu Y, Zhang C, Wong DWK, Liu J, et al. Disc-Aware Ensemble Network for Glaucoma Screening From Fundus Image. IEEE Trans Med Imaging. 2018;37(11):2493-501.
Raghavendra U, Fujita H, Bhandary SV, Gudigar A, Tan JH, Acharya UR. Deep convolution neural network for accurate diagnosis of glaucoma using digital fundus images. Information Sciences. 2018;441:41-9.
Christopher M, Belghith A, Bowd C, Proudfoot JA, Goldbaum MH, Weinreb RN, et al. Performance of Deep Learning Architectures and Transfer Learning for Detecting Glaucomatous Optic Neuropathy in Fundus Photographs. Sci Rep. 2018;8(1):16685.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Enhancing Eye Diseases Diagnosis through Transfer Learning: Study of Deep Convolutional Neural Networks for Accurate Classification of Glaucoma and Diabetic Retinopathy from Healthy Eye Using Fundus Images

Status:

Version 1

Abstract

Figures

Introduction

Methods

Datasets

Kim's Eye Hospital

Drishti-GS1 dataset

DR HAGIS database

APTOS 2019 Blindness Detection

ACRIMA Database

Diabetic Retinopathy Detection 2015

Study design and workflow

Image Preprocessing:

Model Architecture for Transfer Learning:

Experimental Setups, Model Development, and Performance Measurement:

Results

Discussion

Novel Contributions

State of the art

Weaknesses

Conclusion

Declarations

Author contributions

Competing interests

References

Additional Declarations

Status:

Version 1