Hybrid Deep Learning Model for Improved Glaucoma Diagnostic Accuracy

Abstract

Glaucoma is an irreversible neurodegenerative disease that affects the optic nerve, leading to partial or complete vision loss. Early and accurate detection is crucial to prevent vision impairment, which necessitates the development of highly precise diagnostic tools. Deep learning (DL) has emerged as a promising approach for glaucoma diagnosis, where the model is trained on datasets of fundus images. To improve the detection accuracy, we propose a hybrid model for glaucoma detection that combines multiple DL models with two fine-tuning strategies and uses a majority voting scheme to determine the final prediction. In experiments, the hybrid model achieved a detection accuracy of 96.55%, a sensitivity of 98.84%, and a specificity of 94.32%. Integrating datasets was found to improve the performance compared to using them separately even with transfer learning. When compared to individual DL models, the hybrid model achieved a 20.69% improvement in accuracy compared to the best model when applied to a single dataset, a 13.22% improvement when applied with transfer learning across all datasets, and a 1.72% improvement when applied to all datasets. These results demonstrate the potential of hybrid DL models to detect glaucoma more accurately than individual models.

1. Introduction

Glaucoma is a chronic and progressive neuropathy that affects the optic nerve and leads to gradual and irreversible vision loss [1]. By 2040, more than 11 million cases of glaucoma are projected worldwide [2,3]. Interventions such as medication or surgery can significantly improve the quality of life and delay disease progression, but they rarely alter the long-term prognosis [4]. Because glaucoma can be treated in its early stages, timely diagnosis can help prevent blindness [3,5]. Traditionally, glaucoma is diagnosed through clinical tests, such as ophthalmoscopy, fundus photography, nerve fiber layer evaluation, pachymetry, and tonometry [6]. These procedures rely heavily on the expertise of ophthalmologists, are subject to inter-observer variability, and require expensive, non-portable equipment and specialized infrastructure, which limits their use in resource-constrained settings. Furthermore, their diagnostic performance remains moderate, with reported average sensitivity around 70% [7]. Therefore, alternative diagnostic methods are needed that are rapid, accessible, accurate, and low in cost. Automated systems for glaucoma diagnosis have been developed where machine learning (ML) [8,9] or deep learning (DL) [10,11] models are trained on fundus images and have achieved high levels of accuracy, often surpassing the performance of traditional clinical tests. However, such systems still require improvement because of the number of DL models and preprocessing techniques that have not yet been fully evaluated. The limitations of DL models have also been observed in other applications [12], such as bruising-date estimation [13], diabetic retinopathy [14,15], and lung cancer diagnosis [16]. A promising approach to improving the performance of ML and DL models is to use a hybrid model, which generally outperforms individual models [17].

We propose a hybrid model for glaucoma diagnosis that integrates multiple DL models with two fine-tuning strategies and implements a majority voting scheme to determine the final prediction. The fine-tuning strategies were designed to optimize the model performance across different datasets. The proposed hybrid model was implemented using Keras and TensorFlow and was validated against five public datasets with a total of 1707 fundus images.

This research builds upon the results presented in the bachelor’s theses of Nahum Flores and José La Rosa, defended at the National University of San Marcos (UNMSM) in 2022. The current study extends their preliminary work by integrating clinical validation, optimizing the model architecture, and deploying a cross-platform application for real-time glaucoma screening.

2. Related Work

Artificial intelligence (AI) is the field of computer science dedicated to creating systems that simulate human intelligence [18,19]. ML is a branch of AI in which a computer system learns to perform a task based on patterns obtained from input data without being explicitly programed [20,21]. In medical applications, these tasks can include diagnosing diseases such as glaucoma. Figure 1 shows the workflow used by most ML models for glaucoma diagnosis: preprocessing, feature extraction, feature selection, and classification [9,22,23,24,25,26,27,28,29,30,31,32].

Preprocessing techniques include median filtering and histogram equalization [22], contrast-limited adaptive histogram equalization (CLAHE) [24], and region of interest (ROI) cropping [28]. Feature extraction methods include wavelet transform [29,30], GIST descriptors, which are global scene representations that capture the overall spatial layout and texture of an image by summarizing orientation and frequency information through Gabor filter responses [24], and local binary pattern (LBP) [27]. Feature selection methods include genetic algorithms [22], principal component analysis, and Student’s t-test. Classification methods include support vector machines (SVMs) [22,23,24,25,26,27], random forests [28], and decision trees [31].

DL is a branch of ML that incorporates computational models and algorithms to mimic the biological structure of the brain [33]. A specific type of DL is the convolutional neural network (CNN) [34], which comprises a series of processing layers resembling the electrophysiological processes of the visual cortex, where each cortical neuron responds to a specific part of the visual receptive field. Similarly, an artificial neuron or node responds to a specific element of the input data [35]. As shown in Figure 2, a multilayer perceptron (MLP), can be divided into an input layer, an output layer, and several hidden layers. The hidden layers typically comprise convolutional, pooling, and fully connected layers, which are more fully described in Figure 3 [17].

The convolutional layer is the fundamental component of a CNN because it transforms the input data by applying a set of filters (also called kernels) that serve as feature detectors. This filter slides over the input image to produce a feature map as the output, which then undergoes an additional operation called an activation function (e.g., ReLU or softmax) after each convolution operation (Figure 3a) [17,34]. The pooling layer is used to reduce dimensionality and retain the most important high-level features via max pooling (Figure 3b), min pooling (Figure 3c), or average pooling (Figure 3d) [36]. Fully connected layers (Figure 3e) use the high-level features to classify the input image into several classes based on the training dataset (Figure 3f) [21]. After classification, backpropagation is performed to compute the network weights, and this process is repeated many times until the network eventually learns the optimal weights for each node to improve the classification performance [37]. However, the CNN may “memorize” the training data, which means that it cannot be generalized to unseen data. This issue can be addressed by using the early stopping algorithm [38] or dropout algorithms [39].

DL models include AlexNet with five convolutional layers [40]; VGGnet with 16–19 convolutional layers; Inception V1–V4 with 27 convolutional layers; ResNet with 18, 50, and 125 convolutional layers; and DenseNet with 40, 100, 121, and 169 convolutional layers [17]. The performance of a DL model is generally evaluated by using the ImageNet dataset, which comprises more than 1000 categories and 1.2 million images [41]. Compared to AlexNet, recent DL models have achieved better performances owing to unique features such as additional layers, smaller convolutional filters, skip connections, and more complex filters [17,42].

Table 1. DL models for glaucoma diagnosis.

Source	Dataset	Model	Results
[43]	1542 images Kim Eye Hospital (Corea del Sur)	Transfer Learning Inception v3	Accuracy: 84.5% AUC: 0.93
[44]	1426 images Kasturba Medical College (India)	18-layer CNN	Accuracy: 98.13% Sensitivity: 98.0% Specificity: 98.30%
[45]	2554 images Kasturba Medical College (India)	Multi-branch neural network	Accuracy: 91.51% Sensitivity: 92.33% Specificity: 90.90%
[46]	1707 images ACRIMA, HRF, Drishti-GS1, RIM-ONE and sjchoi86-HRF	Fine tuning VGG19	Accuracy: 90.69% Sensitivity: 92.40% Specificity: 88.46%
[47]	8433 images Brussels, Leuven, and Bruges hospitals (Belgium)	Transfer learning ResNet-50	Sensitivity: 95.6% Specificity: 92.0% AUC: 0.986
[48]	10,463 images Beijing Tongren Eye Center	TIA-Net	Accuracy: 85.70% Sensitivity: 84.90% Specificity: 86.9%
[11]	1310 images ACRIMA and ORIGA	Vanilla CNN	Accuracy: 89% Sensitivity: 88%
[49]	113,893 images EyePACS AIROGS	AlterNet-K	Accuracy: 91.6% Sensitivity: 90.7% AUC: 0.9668

summarizes various studies that have applied DL to glaucoma diagnosis and their results. Ahn et al. [43] evaluated a transfer-learned Inception v3 on 1542 fundus images from Kim Eye Hospital (South Korea), reporting 84.5% accuracy and an AUC of 0.93. Raghavendra et al. [44] used an 18-layer CNN trained on 1426 images from Kasturba Medical College (India), achieving 98.13% accuracy, 98.0% sensitivity, and 98.3% specificity. Chai et al. [45] implemented a multi-branch neural network on 2554 images from the same institution, obtaining 91.51% accuracy, 92.33% sensitivity, and 90.90% specificity. Díaz-Pinto et al. [46] fine-tuned VGG19 using 1707 images from ACRIMA, HRF, Drishti-GS1, RIM-ONE, and sjchoi86-HRF, reaching 90.69% accuracy, 92.40% sensitivity, and 88.46% specificity. Hemelings et al. [47] combined transfer and active learning with ResNet-50 on 8433 images from Belgian hospitals, obtaining 95.6% sensitivity, 92.0% specificity, and an AUC of 0.986. Xu et al. [48] proposed TIA-Net, an attention-based model trained on 10,463 images from Beijing Tongren Eye Center, and reported 85.70% accuracy, 84.90% sensitivity, and 86.9% specificity. Ubochi et al. [11] compared a vanilla CNN trained on 1310 images from ACRIMA and ORIGA, achieving 89.0% accuracy and 88.0% sensitivity. Finally, D’Souza et al. [49] introduced AlterNet-K, a compact ResNet + attention model trained on 113,893 EyePACS-AIROGS images, reporting 91.6 % accuracy, 90.7 % sensitivity, and an AUC of 0.968. These works demonstrate that transfer learning, attention mechanisms, and domain-knowledge fusion are key drivers in DL-based glaucoma screening.

3. Materials and Methods

Figure 4 shows the workflow used to develop and validate the proposed hybrid model, which had eight steps. We integrated several processes that have been used in various studies to diagnose glaucoma and other diseases [43,44,45,46,47].

3.1. Dataset Integration

DL models have been shown to perform better with larger datasets [17,33,34]. Therefore, we integrated multiple glaucoma datasets to enhance the diagnostic accuracy. In addition, the proportions of each dataset allocated to training and testing were set to ensure the robustness of the validation results.

Table 2. Datasets used for glaucoma diagnosis.

Dataset	Total	Glaucoma	Normal	Resolution (Pixels)
HRF	45	27	18	3504 × 2336
Drishti-GS1	101	70	31	2896 × 1944
sjchoi86-HRF	401	101	300	2592 × 1728, 2464 × 1632, 1848 × 1224 *
RIM-ONE	455	194	261	478 × 436, 2144 × 1424 *
ACRIMA	705	396	309	2048 × 1536
Total	1707	788	919	-

* Dataset contains images with more than one resolution size.

lists the five public glaucoma datasets that were utilized in this study and are commonly referenced in the literature: HRF [50], Drishti-GS1 [51], sjchoi86-HRF [52], RIM-ONE [53], and ACRIMA [11,46]. Collectively, these datasets contain 1707 fundus images, of which 788 are normal and 919 are glaucomatous.

3.2. DL Model Identification and Fine Tuning

The proposed hybrid model was built by integrating multiple DL models and fine-tuning strategies that have demonstrated strong results in the literature and are widely available to streamline implementation. We considered 12 pretrained DL models: three in the literature (Table 1) and nine additional models available through the Keras API [42], all of which were originally trained on the ImageNet dataset. To optimize these models for the glaucoma diagnosis, we applied two fine-tuning strategies based on the work by Diaz and Pinto [46], resulting in 24 variations. Figure 5 shows the two strategies, which were selected because of their proven effectiveness at enhancing the medical image classification performance of DL models. Strategy 1 (Figure 5a) has a global average pooling layer (GlobalAveragePooling2D) followed by a fully connected layer with two output nodes (glaucoma and normal) using a softmax activation function (AF). Strategy 2 (Figure 5b) has a more complex structure: a flatten layer is followed by a dense layer with 256 nodes using a ReLU AF, a dropout layer, another dense layer with 128 nodes, an additional dropout layer, and a fully connected layer with two output nodes using a softmax AF.

Table 3. Models and Keras layers with fine-tuning strategies.

Model	Original Layers	w/Strategy 1	w/Strategy 2
DenseNet121	427	429	433
DenseNet169	595	597	601
InceptionV3	311	313	317
MobileNet	86	88	92
MobileNetV2	154	156	160
NASNetMobile	769	771	775
ResNet101	345	347	351
ResNet50	175	177	181
ResNet50V2	190	192	196
VGG16	19	21	25
VGG19	22	24	28

lists the Keras layers of the selected models with each strategy.

3.3. Hybrid Model

The classification performance for medical images can be improved by combining multiple DL models [17]. In this study, we employed a majority voting strategy with equal weights rather than a weighted scheme, based on practical and clinical considerations (Figure 6). First, this approach is easier to deploy and maintain across healthcare institutions, as it avoids the need to recalibrate voting weights whenever the dataset is updated or expanded, simplifying the integration of new data [54]. Second, majority voting offers greater transparency: each final decision can be traced to a simple consensus among individual models, which aligns with common clinical decision-making practices and facilitates explainability through well-established methods, such as SHAPs (SHapley Additive exPlanations) and PDPs (Partial Dependence Plots), which help reveal how input features influence model predictions [55]. Furthermore, evidence from ophthalmic imaging supports the effectiveness of this strategy. In a study on scanning laser ophthalmoscopy images, majority voting achieved significantly better classification results than weighted voting, particularly in scenarios with limited training data [56]. Similarly, in a glaucoma screening task using retinal images, an ensemble model based on equal-probability soft voting improved sensitivity without sacrificing specificity [57]. Based on these findings, we selected majority voting for its balance between performance, simplicity, and interpretability in the context of automated glaucoma detection.

3.4. Preprocessing

Preprocessing is important for improving the performance of DL models because a high image resolution directly affects the training time [58,59,60]. Preprocessing steps include cropping a fundus image centered on the optic disk, resizing the image, and data augmentation. As shown in Figure 7, the HRF [50], Drishti-GS1 [51], and sjchoi86-HRF [52] datasets, were manually cropped by ophthalmologists, using the optic disk radius to extract a square region with a side length three times that value. In contrast, the RIM-ONE [53] and ACRIMA [11,46] datasets already consisted of cropped optic disk images and were used as provided. All images were then resized to match the default input size of each model, as given in

Table 4. Input image sizes for each model.

Model	Default Input Shape Size (Pixels)
DenseNet121	224 × 224
DenseNet169	224 × 224
MobileNet	224 × 224
MobileNetV2	224 × 224
ResNet101	224 × 224
ResNet50	224 × 224
ResNet50V2	224 × 224
VGG16	224 × 224
VGG19	224 × 224
InceptionV3	299 × 299
Xception	299 × 299
NASNetMobile	331 × 331

. Performance enhancement techniques such as CLAHE were not applied because they do not guarantee an improved classification performance [61]. As shown in Figure 8, we also used data augmentation to increase the number of images and reduce overfitting. The applied transformations included horizontal flip, vertical flip, combined horizontal and vertical flip, 30° rotation, and 1.2× magnification. These transformations were applied while maintaining their labels [62].

3.5. Performance Metrics

Commonly used performance metrics in medical image analysis include the sensitivity (Sen), specificity (Esp), and accuracy (Acc) [63]. In this study, we defined a true positive (TP) as a glaucomatous image that is correctly classified, a true negative (TN) as a normal image that is correctly classified, a false negative (FN) as a glaucomatous image that is classified as normal, and a false positive (FP) as a normal image that is classified as glaucomatous. Thus, we defined the sensitivity as the percentage of images correctly classified as glaucomatous:

$$\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}\times 100\mathrm{\%}$$

(1)

We defined the specificity as the percentage of images correctly classified as normal:

$$\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}}}\times 100\mathrm{\%}$$

(2)

Finally, we defined the accuracy as the percentage of correctly classified images:

$$\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}}}\times 100\mathrm{\%}$$

(3)

To ensure the reliability of the performance evaluation, we used the k-fold cross-validation technique (k = 10) [46].

3.6. Training and Testing

Figure 9 shows our evaluation scheme for training, validating, and testing each DL model. We followed a systematic approach to execute multiple experiments and achieve optimal results. Each experiment corresponded to a different configuration of model hyperparameters, which are given in

Table 5. Training parameters for each scenario.

Scenario	Dataset	Epochs	Learning Rate	Batch Size
1	HRF	100	1 × 10−4	1
	Drishti-GS1	100	1 × 10−4	1
	sjchoi86-HRF	100	1 × 10−4	8
	RIM-ONE	100	1 × 10−4	8
	ACRIMA	50	1 × 10−4	8
2	choi + HRF	200	1 × 10−4	8
	Drishti + RIM	200	1 × 10−4	8
	Acrima	50	1 × 10−4	8
3	All Datasets	200	1 × 10−4	8

. Different training scenarios were formulated to evaluate the impacts of using multiple datasets, transfer learning, and other factors. The evaluation scheme was applied to each selected model with both fine-tuning strategies (Table 3) using the model hyperparameters (Table 5) for three different scenarios, which resulted in 216 experiments in total.

Figure 10 shows the three training scenarios tested in this study to analyze the impact of different factors. Scenario 1 (Figure 10a) analyzed the impact of different and larger testing sets. Each model was trained with 90% of a dataset, after which two tests were performed. Test 1 measured the effectiveness of the trained models when applied to the unused data (i.e., remaining 10%) of the same dataset. Test 2 measured the effectiveness of the trained models when applied to the unused data (i.e., remaining 10%) of all five datasets. Scenario 2 (Figure 10b) analyzed the impact of transfer learning through a successive fine-tuning strategy. First, each model was trained on a combined set consisting of 90% of the HRF and sjchoi86-HRF datasets. Using the resulting weights, the model was then fine-tuned on a new combined set comprising 90% of the Drishti-GS1 and RIM-ONE datasets. Finally, transfer learning was applied once more by fine-tuning the model on 90% of the ACRIMA dataset. This progressive approach allowed the model to gradually adapt to increasingly heterogeneous data. After the final fine-tuning stage, all models were evaluated using the remaining 10% of all five datasets. Scenario 3 (Figure 10c) analyzed the impact of combining multiple datasets. Each model was trained on a combined set integrating 90% of all five datasets and was then tested on the remaining 10% of data.

3.7. Implementation

DL models are commonly built using modern frameworks such as TensorFlow, PyTorch, and Keras, which offer high-level APIs and GPU support for efficient training. The choice of platform often depends on deployment simplicity, community support, and integration with existing infrastructure. Additionally, it is important to consider whether the hardware is local or cloud-based. For example, a cloud-based implementation could allow for the inclusion of an intelligent system that can be accessed via the Web or a mobile device from anywhere with Internet access. In this study, we implemented the DL models by using Python 3.6, Keras 2.4.3, and TensorFlow 2.4.0 on Google Colab [64] with a 2.3 GHz Xeon CPU, 13 GB RAM, and 16 GB NVIDIA Tesla V100 GPU. The total execution time for all experiments was around 350 h. As shown in Figure 11, we also used IONIC 8 and Django 4.2 to develop a cross-platform application that allows users to obtain a diagnosis by uploading a fundus image. Our model, along with the source code of the cross-platform application and the links to download the five public datasets, are available in the following GitHub repository: https://github.com/NahumFGz/DLG-C22200401 (accessed on 1 July 2025). This repository includes Python scripts used for training and evaluation, configuration files, the IONIC front-end code, and the Django back-end implementation. Detailed instructions are provided to enable full reproducibility of the experiments.

3.8. Results

The models were evaluated according to the performance metrics for each scenario. For clarity, we used the notation ModelName_m1 and ModelName_m2 to distinguish between the same model using fine-tuning strategies 1 and 2, respectively (Figure 5). The results of the individual models were then compared against the results of the hybrid model implementing the majority voting scheme. The notation voting N means that the hybrid model incorporated N models with the best accuracy among the 24 models considered.

4. Results and Discussion

4.1. Effects of Training Scenarios

Table 6. Results for scenario 1 in test 1.

N°	Model	HRF (%)			Drishti-GS1 (%)			Sjchoi86-HRF (%)			Rim-One (%)			Acrima (%)
		Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp
1	DenseNet121_m1	80.00	100.00	66.67	72.73	100.00	0.00	85.37	80.00	87.10	84.78	95.65	73.91	98.59	97.67	100.00
2	DenseNet121_m2	60.00	100.00	33.33	72.73	100.00	0.00	90.24	90.00	90.32	89.13	91.30	86.96	98.59	100.00	96.43
3	DenseNet169_m1	60.00	50.00	66.67	54.55	62.50	33.33	80.49	90.00	77.42	91.30	95.65	86.96	97.18	95.35	100.00
4	DenseNet169_m2	60.00	100.00	33.33	72.73	87.50	33.33	87.80	70.00	93.55	93.48	91.30	95.65	94.37	90.70	100.00
5	InceptionV3_m1	60.00	50.00	66.67	72.73	87.50	33.33	90.24	80.00	93.55	86.96	91.30	82.61	97.18	95.35	100.00
6	InceptionV3_m2	60.00	100.00	33.33	72.73	75.00	66.67	90.24	70.00	96.77	93.48	95.65	91.30	97.18	95.35	100.00
7	MobileNet_m1	60.00	100.00	33.33	63.64	62.50	66.67	90.24	80.00	93.55	89.13	86.96	91.30	97.18	97.67	96.43
8	MobileNet_m2	60.00	100.00	33.33	90.91	100.00	66.67	85.37	70.00	90.32	93.48	95.65	91.30	95.77	93.02	100.00
9	MobileNetV2_m1	60.00	0.00	100.00	54.55	37.50	100.00	82.93	70.00	87.10	95.65	95.65	95.65	80.28	100.00	50.00
10	MobileNetV2_m2	60.00	50.00	66.67	72.73	100.00	0.00	85.37	90.00	83.87	91.30	95.65	86.96	64.79	100.00	10.71
11	NASNetMobile_m1	60.00	100.00	33.33	54.55	75.00	0.00	80.49	40.00	93.55	71.74	91.30	52.17	63.38	100.00	7.14
12	NASNetMobile_m2	60.00	0.00	100.00	72.73	87.50	33.33	82.93	40.00	96.77	76.09	60.87	91.30	63.38	72.09	50.00
13	ResNet101_m1	80.00	50.00	100.00	54.55	50.00	66.67	85.37	90.00	83.87	82.61	82.61	82.61	94.37	90.70	100.00
14	ResNet101_m2	60.00	100.00	33.33	72.73	75.00	66.67	82.93	80.00	83.87	91.30	91.30	91.30	95.77	97.67	92.86
15	ResNet50_m1	80.00	50.00	100.00	54.55	50.00	66.67	87.80	70.00	93.55	95.65	95.65	95.65	95.77	97.67	92.86
16	ResNet50_m2	60.00	100.00	33.33	72.73	75.00	66.67	85.37	50.00	96.77	95.65	95.65	95.65	97.18	97.67	96.43
17	ResNet50V2_m1	80.00	100.00	66.67	54.55	62.50	33.33	78.05	60.00	83.87	84.78	86.96	82.61	76.06	60.47	100.00
18	ResNet50V2_m2	80.00	100.00	66.67	63.64	75.00	33.33	85.37	70.00	90.32	89.13	91.30	86.96	91.55	90.70	92.86
19	VGG16_m1	60.00	0.00	100.00	81.82	100.00	33.33	90.24	90.00	90.32	82.61	95.65	69.57	91.55	100.00	78.57
20	VGG16_m2	40.00	100.00	0.00	72.73	100.00	0.00	90.24	90.00	90.32	91.30	91.30	91.30	92.96	95.35	89.29
21	VGG19_m1	60.00	50.00	66.67	72.73	100.00	0.00	87.80	70.00	93.55	89.13	91.30	86.96	97.18	97.67	96.43
22	VGG19_m2	40.00	100.00	0.00	72.73	100.00	0.00	90.24	70.00	96.77	93.48	95.65	91.30	97.18	97.67	96.43
23	Xception_m1	60.00	50.00	66.67	72.73	100.00	0.00	90.24	70.00	96.77	93.48	91.30	95.65	92.96	90.70	96.43
24	Xception_m2	60.00	50.00	66.67	81.82	87.50	66.67	87.80	70.00	93.55	89.13	95.65	82.61	95.77	95.35	96.43
25	Average	62.50	70.83	56.94	68.94	81.25	36.11	86.38	72.92	90.73	88.95	91.30	86.59	90.26	93.70	84.97

and

Table 7. Results for scenario 1 in test 2.

N°	Model	HRF (%)			Drishti-GS1 (%)			sjchoi86-HRF (%)			Rim-One (%)			Acrima (%)
		Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp
1	DenseNet121_m1	58.62	70.93	46.59	54.02	91.86	17.05	64.94	39.53	89.77	66.09	98.84	34.09	70.11	96.51	44.32
2	DenseNet121_m2	68.39	59.30	77.27	50.00	97.67	3.41	64.37	56.98	71.59	66.09	91.86	40.91	75.86	94.19	57.95
3	DenseNet169_m1	49.43	91.86	7.95	58.05	24.42	90.91	71.26	66.28	76.14	72.41	93.02	52.27	71.84	95.35	48.86
4	DenseNet169_m2	71.84	82.56	61.36	59.20	65.12	53.41	63.79	29.07	97.73	75.86	88.37	63.64	71.26	72.09	70.45
5	InceptionV3_m1	50.57	23.26	77.27	55.75	59.30	52.27	65.52	34.88	95.45	66.09	94.19	38.64	64.37	97.67	31.82
6	InceptionV3_m2	51.15	19.77	81.82	58.62	79.07	38.64	66.09	39.53	92.05	78.16	79.07	77.27	68.39	100.00	37.50
7	MobileNet_m1	49.43	98.84	1.14	52.87	37.21	68.18	64.94	36.05	93.18	67.24	66.28	68.18	66.09	95.35	37.50
8	MobileNet_m2	64.37	74.42	54.55	62.64	53.49	71.59	68.39	43.02	93.18	72.99	70.93	75.00	68.39	94.19	43.18
9	MobileNetV2_m1	51.72	5.81	96.59	52.87	4.65	100.00	59.77	24.42	94.32	68.39	59.30	77.27	59.20	100.00	19.32
10	MobileNetV2_m2	59.77	44.19	75.00	52.30	95.35	10.23	58.62	27.91	88.64	71.26	84.88	57.95	52.30	97.67	7.95
11	NASNetMobile_m1	47.70	87.21	9.09	52.87	75.58	30.68	56.32	15.12	96.59	65.52	69.77	61.36	49.43	96.51	3.41
12	NASNetMobile_m2	49.43	100.00	0.00	57.47	19.77	94.32	57.47	17.44	96.59	62.07	72.09	52.27	48.85	63.95	34.09
13	ResNet101_m1	51.15	1.16	100.00	52.87	13.95	90.91	75.29	66.28	84.09	59.77	63.95	55.68	64.37	95.35	34.09
14	ResNet101_m2	57.47	30.23	84.09	64.37	84.88	44.32	74.14	63.95	84.09	72.41	89.53	55.68	67.82	100.00	36.36
15	ResNet50_m1	58.62	37.21	79.55	55.75	11.63	98.86	75.29	63.95	86.36	67.24	75.58	59.09	63.79	98.84	29.55
16	ResNet50_m2	58.62	44.19	72.73	68.97	55.81	81.82	67.82	48.84	86.36	66.67	83.72	50.00	64.94	97.67	32.95
17	ResNet50V2_m1	50.00	63.95	36.36	56.32	37.21	75.00	61.49	40.70	81.82	72.99	81.40	64.77	72.41	97.67	47.73
18	ResNet50V2_m2	57.47	70.93	44.32	70.69	60.47	80.68	64.94	30.23	98.86	71.26	79.07	63.64	70.11	90.70	50.00
19	VGG16_m1	55.17	9.30	100.00	70.69	48.84	92.05	73.56	54.65	92.05	64.37	95.35	34.09	62.07	100.00	25.00
20	VGG16_m2	49.43	100.00	0.00	49.43	100.00	0.00	50.57	0.00	100.00	72.41	86.05	59.09	62.07	90.70	34.09
21	VGG19_m1	60.92	67.44	54.55	69.54	63.95	75.00	74.14	58.14	89.77	73.56	84.88	62.50	64.37	96.51	32.95
22	VGG19_m2	49.43	100.00	0.00	49.43	100.00	0.00	69.54	45.35	93.18	64.94	90.70	39.77	64.94	97.67	32.95
23	Xception_m1	50.00	98.84	2.27	50.00	100.00	1.14	63.22	31.40	94.32	67.82	77.91	57.95	72.41	90.70	54.55
24	Xception_m2	48.28	19.77	76.14	63.79	56.98	70.45	68.39	44.19	92.05	60.92	83.72	38.64	66.09	91.86	40.91
25	Average	54.96	58.38	51.61	57.85	59.88	55.87	65.83	40.75	90.34	68.61	81.69	55.82	65.06	93.80	36.98
26	Voting 3	64.9	86.05	44.32	71.3	53.49	88.64	76.4	65.1	87.5	77.01	87.21	67.05	72.4	96.51	48.86
27	Voting 5	68.4	77.91	59.09	74.7	60.47	88.64	75.9	58.1	93.18	77.59	84.88	70.45	73	95.35	51.14
28	Voting 7	69.5	73.26	65.91	71.3	61.63	80.68	75.9	59.3	92.05	75.86	90.70	61.36	70.7	95.35	46.59
29	Voting 9	69.5	63.95	75	70.7	65.12	76.14	71.8	50	93.18	74.71	89.53	60.23	70.1	97.67	43.18
30	Voting 11	67.8	54.65	80.68	70.1	56.98	82.95	71.8	50	93.18	74.71	87.21	62.50	70.1	100	40.91

present the results for training with separate datasets (scenario 1). Across the three performance metrics, the performance of a DL model trained on a single dataset decreased by an average of 16.94% from test 1 (using the remaining data from the same dataset) to test 2 (using data from other datasets). Thus, DL models trained on a single dataset were less reliable than those trained on multiple datasets. The performance of the DL models improved with larger and more heterogeneous datasets.

Table 8. Results for scenarios 2 and 3.

N°	Model	Scenario 2 (%)			Scenario 3 (%)
		Acc	Sen	Esp	Acc	Sen	Esp
1	DenseNet121_m1	77.01	100.00	54.55	90.80	82.56	98.86
2	DenseNet121_m2	74.14	96.51	52.27	93.10	90.70	95.45
3	DenseNet169_m1	79.31	95.35	63.64	92.53	93.02	92.05
4	DenseNet169_m2	77.01	97.67	56.82	93.68	94.19	93.18
5	InceptionV3_m1	70.69	96.51	45.45	92.53	95.35	89.77
6	InceptionV3_m2	82.76	95.35	70.45	93.68	95.35	92.05
7	MobileNet_m1	75.86	95.35	56.82	93.68	98.84	88.64
8	MobileNet_m2	75.86	93.02	59.09	90.23	90.70	89.77
9	MobileNetV2_m1	72.41	82.56	62.50	93.68	95.35	92.05
10	MobileNetV2_m2	82.18	75.58	88.64	91.95	96.51	87.50
11	NASNetMobile_m1	76.44	90.70	62.50	91.95	93.02	90.91
12	NASNetMobile_m2	62.64	97.67	28.41	90.23	91.86	88.64
13	ResNet101_m1	71.26	100.00	43.18	91.38	95.35	87.50
14	ResNet101_m2	79.31	93.02	65.91	92.53	93.02	92.05
15	ResNet50_m1	68.97	96.51	42.05	94.83	98.84	90.91
16	ResNet50_m2	74.14	95.35	53.41	94.25	96.51	92.05
17	ResNet50V2_m1	83.33	87.21	79.55	91.38	94.19	88.64
18	ResNet50V2_m2	77.59	97.67	57.95	94.25	93.02	95.45
19	VGG16_m1	65.52	91.86	39.77	90.23	87.21	93.18
20	VGG16_m2	66.09	91.86	40.91	94.25	94.19	94.32
21	VGG19_m1	69.54	96.51	43.18	91.95	91.86	92.05
22	VGG19_m2	66.09	94.19	38.64	90.80	91.86	89.77
23	Xception_m1	78.74	96.51	61.36	91.95	93.02	90.91
24	Xception_m2	77.01	96.51	57.95	93.10	93.02	93.18
25	Average	74.33	93.89	55.21	92.46	93.31	91.62

presents the results for scenarios 2 and 3. A comparison using the ACRIMA dataset showed that transfer learning (scenario 2) improved the results of each DL model by an average of 9.27% compared with scenario 1 (test 2 (Table 7)). This result is consistent with other studies [12,43,47]. Therefore, transfer learning is recommended to improve the training performance of DL models.

Finally, training with an integrated dataset (scenario 3) yielded better results than training with separate datasets (scenario 1) or transfer learning (scenario 2) with improvements of 23.85% and 18.13%, respectively. This result aligns with several studies showing that larger datasets lead to better results [17,33,34,46]. Moreover, greater heterogeneity (i.e., multiple datasets) results in better generalization and thus better outcomes. Therefore, integrating multiple datasets is recommended for DL models.

4.2. Performance of the Hybrid Model

The results for the different training scenarios (Table 6, Table 7 and Table 8) showed that the best results were obtained by using an integrated dataset (i.e., scenario 3). In this scenario, the best DL model was ResNet50 with fine-tuning strategy 1, which achieved an accuracy of 94.83%, a sensitivity of 98.84%, and a specificity of 90.91%.

Table 9. Results of the hybrid model according to the number of voting models.

N°	Model	Scenario 1 (%)			Scenario 2 (%)			Scenario 3 (%)
		Acc	Sen	Esp	Acc	Sen	Esp	Acc	Sen	Esp
1	Voting 3	77.01	87.21	67.05	88.51	90.70	86.36	96.55	98.84	94.32
2	Voting 5	77.59	84.88	70.45	84.48	94.19	75.00	96.55	97.67	95.45
3	Voting 7	75.86	90.70	61.36	84.48	96.51	72.73	95.98	97.67	94.32
4	Voting 9	74.71	89.53	60.23	81.61	98.84	64.77	95.98	97.67	94.32
5	Voting 11	74.71	87.21	62.50	82.76	100.00	65.91	95.98	97.67	94.32

presents the results for the hybrid model. It yielded better results than individual DL models across all scenarios, and the best results were achieved with Voting 3 and Voting 5. Voting 3 (i.e., ResNet50 with fine-tuning strategies 1 and 2 and ResNet50V2 with fine-tuning strategy 2) achieved an accuracy of 96.55% in scenario 3, which outperformed the best DL model in scenario 1 (DenseNet169_m2) by 20.69%, the best DL model in scenario 2 (ResNet50V2_m1) by 13.22%, and best DL model in scenario 3 (ResNet50V2_m1) by 1.72%. The models performed best in scenario 3, and Voting 3 and Voting 5 outperformed the individual models.

5. Conclusions

The results of this study demonstrate the feasibility of applying a hybrid DL model to developing an automated system for accurately diagnosing glaucoma from fundus images. The best-performing hybrid model (Voting 3) achieved an accuracy of 96.55%, a sensitivity of 98.84%, and a specificity of 94.32%, surpassing the performance of human specialists. Implementing this hybrid model in a cloud-based system will facilitate glaucoma diagnosis anywhere with Internet access at a low cost, which will be beneficial for areas with a shortage of specialists. We expect our methodology to also be applicable to diagnosing other ocular diseases based on fundus images.