Enhancing the Performance of Multi-Class Classification Systems for Retinal Diseases Using Deep Learning Models

Retinal diseases are a major cause of visual impairment and blindness worldwide, affecting individuals across all age groups. Visual signals are captured by the retina and transmitted to the brain through the optic nerve [1, 2]. Several conditions, including cataract, glaucoma, diabetic retinopathy, and age-related macular degeneration, can lead to severe and irreversible vision loss if diagnosis or treatment is delayed [3]. The World Health Organization estimates that 2.2 billion people experience vision impairment, and a substantial proportion of cases could be prevented or treated through timely diagnosis and intervention [4]. Cataract and diabetic retinopathy account for a large number of preventable cases, and glaucoma remains a major cause of permanent vision loss [5].

Manual screening of retinal fundus images is labor-intensive and depends on expert interpretation, which may lead to variability and delayed clinical decisions [6, 7]. Recent advances in artificial intelligence have enabled automated analysis of ophthalmic images using deep learning, particularly deep convolutional neural networks (DCNNs) [8-11]. Despite their success, conventional end-to-end deep models often require substantial computational resources and may generalize poorly across datasets, especially when image acquisition conditions differ. In addition, many DCNNs provide limited transparency, which can reduce clinical trust [12].

Ensemble learning is a well-established strategy to improve robustness and generalization by combining multiple models. Methods such as hard voting, soft voting, stacking, bagging, and boosting can reduce variance and mitigate the limitations of individual classifiers [13-15]. Nevertheless, many existing retinal classification studies focus primarily on maximizing accuracy, while offering limited analysis of computational cost and limited interpretability beyond qualitative examples [16].

To address these limitations, this work makes the following contributions:

(1) A hybrid framework is proposed that combines two complementary CNN backbones (InceptionV3 and DenseNet121) with multiple ensemble learning strategies (hard voting, soft voting, stacking, bagging, and gradient boosting) for four-class retinal disease classification.

(2) Single-CNN and ensemble configurations are compared using predictive metrics and training time to assess practical deployability.

(3) Grad-CAM is used to generate visual explanations and to examine whether the learned attention patterns align with clinically meaningful retinal structures.

The overall objective is to deliver a screening system that is accurate, computationally efficient, and interpretable.

Figure 1 illustrates the overall workflow of the proposed approach. The pipeline includes dataset preparation, image preprocessing, deep feature extraction using two CNN backbones, feature fusion, ensemble classification, and interpretability analysis using Grad-CAM.

Figure 1. Proposed scheme

3.1 Dataset overview

The Eye Diseases Classification dataset contains 4,217 color fundus images distributed across four classes: 1,074 normal, 1,038 cataract, 1,007 glaucoma, and 1,098 diabetic retinopathy [27]. The class distribution is approximately balanced. Therefore, no over-sampling, under-sampling, or class-weighting strategy was applied. Instead, a stratified split was used to preserve class proportions.

The dataset was partitioned into 80% training data (3,373 images) and 20% held-out test data (844 images) using stratification at the image level. During training, 10% of the training set was further reserved for validation to support early stopping and hyperparameter selection. All reported results correspond to the held-out test set, which was not used during training or model selection (Table 2).

Table 2. Dataset overview

Figure 2. Retinal images from the dataset

Representative examples from each class are shown in Figure 2. Normal images present typical retinal anatomy without visible pathology. Cataract images often exhibit reduced contrast due to lens opacity. Glaucoma images are characterized by changes around the optic nerve head, including increased cupping. Diabetic retinopathy images may contain microaneurysms, hemorrhages, and exudates.

3.2 Preprocessing

All images were resized to 224 × 224 pixels to ensure consistent input dimensions. Pixel intensities were normalized to the [0,1] range to stabilize training. The images were kept in RGB format because color information is clinically relevant in fundus imaging.

To reduce overfitting and improve robustness, data augmentation was applied during training. The augmentation included random horizontal flips, small rotations (±10°), random zooming (up to 10%), and mild brightness and contrast variations. After preprocessing, the training set contained 3,373 images and the held-out test set contained 844 images [28].

Deep learning-based modeling: In this study, two ImageNet-pretrained CNN architectures were used as feature extractors: InceptionV3 and DenseNet121. For each backbone, the original classification head was removed and replaced by a GAP layer, followed by a fully connected layer with 256 units (ReLU) and dropout (0.5). A final softmax layer with four outputs was used during CNN training to support supervised feature learning. Deep features were extracted from the penultimate layer and later used to train ensemble classifiers.

• Inception

InceptionV3 employs parallel convolutional operations with multiple kernel sizes within Inception modules, allowing multi-scale feature extraction. This design improves representational capacity while controlling computational cost through factorized convolutions and 1 × 1 bottleneck layers. Figure 3 shows the InceptionV3 architecture used in this work [29, 30].

Figure 3. Inception model architecture [29]

• DenseNet

DenseNet121 introduces dense connectivity, where each layer receives feature maps from all preceding layers and passes its own feature maps to all subsequent layers. This design strengthens gradient flow, encourages feature reuse, and reduces redundant computation. DenseNet121 is therefore well suited for medical image classification tasks that require fine-grained feature representation. Figure 4 shows the DenseNet121 architecture [31].

Figure 4. DenseNet121 model architecture [32]

The DenseNet121 is a DCNN that improves feature information sharing and gradient path with deep network connectivity. DenseNet layers connect to every one of the preceding layers, accepting their feature maps for processing until they propagate the results to all subsequent layers, which reduces processing redundancy [32]. The model design utilizes four dense blocks with a layer and includes transition layers containing 1 × 1 convolutions together with 2 × 2 average pooling elements. The first step combines a 7 × 7 convolution filter with a 3 × 3 max pooling operation to handle the input before the dense blocks receive it.

The distribution of 3 × 3 convolutional layers with BatchNorm and ReLU activation functions enables each dense block to improve stability and convergence. The model includes dropout layers, which help prevent overfitting, while global average pooling (GAP) reduces the feature dimensions before its output gets sent to a fully connected (FC) layer with a softmax classifier for prediction purposes. The results showed DenseNet121 as an efficient and better feature propagation, so it’s an optimal choice for medical image classification, giving high accuracy and computational efficiency performance.

3.3 Ensemble learning

To improve classification robustness and generalization, ensemble learning methods were applied to the fused deep features extracted from InceptionV3 and DenseNet121. Ensemble learning combines multiple predictors to reduce variance and improve stability, which is particularly beneficial in medical imaging tasks. The following ensemble strategies were evaluated.

• Hard Voting

Hard voting assigns the final class label based on the majority vote across base classifiers [33]. Each classifier produces a discrete class prediction, and the class with the highest vote count is selected.

• Soft Voting

Soft voting aggregates predicted class probabilities instead of discrete labels. The probability distributions produced by base models are averaged, and the class with the highest mean probability is selected [34, 35]. Soft voting typically improves performance when probability estimates are well calibrated.

• Stacking

Stacking trains multiple base models and uses their outputs as inputs to a meta-learner. The meta-learner learns how to optimally combine base predictions, often improving generalization compared with simple voting [36, 37].

• Bagging (Random Forest)

Bagging trains multiple models on bootstrap samples of the training data and aggregates their predictions by majority vote [38]. Random Forest is a commonly used bagging model based on decision trees. It reduces variance, is robust to noise, and can model complex decision boundaries.

3.4 Interpretability using Grad-CAM

Grad-CAM was applied to the last convolutional layers of the trained CNNs to generate class-specific heatmaps. For glaucoma, activations were concentrated around the optic nerve head and cup-to-disc region, which are clinically relevant for assessing glaucomatous damage. For diabetic retinopathy, the highlighted regions often corresponded to lesion patterns near the posterior pole, including areas consistent with microaneurysms and exudates. In normal images, activation tended to focus on anatomical landmarks such as the optic disc and major vessels without emphasizing irrelevant background regions.

For cataract images, attention patterns were more diffuse. This behavior is expected because cataract primarily affects the lens and can reduce global image contrast rather than producing localized retinal lesions. Overall, Grad-CAM results suggested that the models relied on clinically plausible cues. However, this analysis remained qualitative, and future work should include expert evaluation to quantify agreement between explanations and clinical criteria [39].

3.5 Training configuration and hyperparameters

InceptionV3 and DenseNet121 were initialized with ImageNet pretrained weights. Input images were resized to 224 × 224 pixels and normalized to [0,1]. Data augmentation was applied online during training.

Both CNNs were trained using the Adam optimizer with an initial learning rate of 1×10⁻⁴, a batch size of 32, and categorical cross-entropy loss. Training was performed for up to 50 epochs. Early stopping monitored validation loss with a patience of 7 epochs, and the learning rate was reduced on plateau (factor 0.5, patience 3).

Ensemble models (Random Forest bagging and gradient boosting) were trained using deep features extracted from the penultimate layer of the CNNs. Hyperparameters such as the number of trees, maximum depth, and learning rate were tuned using grid search on the training and validation sets. All experiments were implemented in Python using TensorFlow/Keras and scikit-learn and were executed on a GPU-enabled workstation.

5.1 Inception results

Table 3 reports class-wise precision, recall, and F1-score for InceptionV3. The model achieved an overall accuracy of 0.85 on the test set. Performance was highest for cataract and diabetic retinopathy, while glaucoma exhibited lower recall, suggesting that some glaucoma cases were misclassified.

Table 3. Classification of the inception model

Figure 5 shows the learning curves for training and validation. Training accuracy increased steadily and approached saturation. Validation accuracy stabilized after the initial epochs, indicating reasonable generalization. Training and validation loss decreased overall, although minor fluctuations were observed in validation loss, which may indicate residual sensitivity to sample variability.

Figure 6 presents the ROC curves. The AUC values were high across classes, reaching 0.99 for cataract and diabetic retinopathy and 0.96 for glaucoma and normal. These results indicate strong separability, although classification errors remained more frequent for glaucoma.

Figure 5. Learning curve of InceptionV3

Figure 6. ROC curve of InceptionV3

5.2 DenseNet results

Table 4 summarizes performance for DenseNet121. The model achieved 0.91 test accuracy. Diabetic retinopathy yielded the highest performance, with an F1-score of 0.99. Cataract also achieved a strong performance, while glaucoma remained the most challenging class due to lower recall (0.81).

Table 4. Classification of DenseNet

Figure 7. ROC curve of DenseNet

Figure 7 shows the ROC curves for DenseNet121. AUC values were 1.00 for diabetic retinopathy, 0.99 for cataract, 0.98 for normal, and 0.97 for glaucoma, confirming strong multi-class discrimination.

5.3 Ensemble results

Table 5 compares ensemble strategies using accuracy, precision, recall, and F1-score. Bagging with Random Forest achieved the highest accuracy (0.9941), indicating strong robustness and stability. Stacking, soft voting, and hard voting also improved performance relative to individual CNN models, although with smaller gains.

Table 5. Comparison analysis of learning techniques used

Table 6 compares training time after applying the training process by the graphics card RTX 4060 Ti and using parallel process for evaluating a short time, also the macro-average accuracy for InceptionV3, DenseNet121, and the combined CNN feature ensemble. DenseNet121 achieved higher macro-average accuracy than InceptionV3 and required shorter training time. The fused feature ensemble further improved macro-average accuracy with a moderate increase in total computation.

Table 6. Training time with the accuracy comparison

5.3.1 Results of best model: Bagging (Random Forest)

Table 7 reports class-wise performance for bagging with Random Forest. Cataract, diabetic retinopathy, and normal achieved perfect precision, recall, and F1-score. Glaucoma achieved near-perfect performance with an F1-score of 0.99. Overall test accuracy reached 0.99, and macro-average metrics remained consistently high across classes.

Table 7. Classification of bagging (Random Forest)

5.3.2 Improve boost results

Table 8 shows performance for the optimized boosting model. All classes achieved precision, recall, and F1-score of 1.00, resulting in 1.00 overall accuracy on the test set. Although this outcome is strong, it should be interpreted cautiously given the single-dataset evaluation and the potential for overfitting.

Table 8. Classification of optimized boosting

5.4 Final comparison

Table 9 summarizes the performance of all evaluated models. Bagging with Random Forest produced the highest accuracy among the evaluated strategies. DenseNet121 outperformed InceptionV3, suggesting that it captured discriminative retinal patterns more effectively in this setting. Overall, ensemble models improved performance over individual CNNs, which supports the effectiveness of combining deep features with classical ensemble learners for retinal disease classification.

Table 9. Comparison results

5.5 Grad-CAM results

Figure 8 shows representative Grad-CAM visualizations. The heatmaps highlight image regions that contributed most to the predicted class. High-importance regions are indicated by warm colors. In several examples, activations focused on clinically relevant areas, such as the optic disc and vascular structures, supporting interpretability of the model decisions.

Figure 8. Grad-CAM results

Strong activation around the optic disc and surrounding vascular structures is shown in the first row of the heat map, indicating that the model uses these features for classification. With a distinct focus in the central retinal region, the activation in the second row is more localized and may suggest pathology. These findings demonstrate that the model successfully detects important diagnostic characteristics, improving the explanatory ability and reliability of AI-assisted medical diagnostics.

5.6 Discussion

Although the proposed ensembles achieved very high test accuracies (up to 99–100%), these results should be interpreted cautiously. First, evaluation was conducted on a single public dataset, which may not reflect the variability observed in real clinical environments, including differences in acquisition devices, illumination, and patient demographics. Second, even with data augmentation, early stopping, and a held-out test set, overfitting cannot be fully excluded, particularly for tree-based models that can fit complex decision boundaries when feature representations are highly separable.

Future work will therefore focus on external validation using additional datasets, cross-dataset testing, and prospective evaluation in clinical workflows. Further analysis should also investigate calibration, error patterns by disease severity, and model robustness under domain shifts.