In image classification, the Deep Learning [1], especially with Convolutional Neural Networks (CNN) enables a model to automatically learn features directly from raw images and further process. The model employs convolutional layers to identify elements such as edges, forms, and textures, which accumulate complex patterns as the layers advance, beginning with an input layer where pixel data is fed into the network. The collected characteristics go through multiple layers before being flattened into a vector and then combined into high-level, meaningful representations by going through fully connected layers. Lastly, class probabilities are provided by the output layer; the anticipated category is indicated by the highest class probability. Deep learning is a potent techniques for a variety of applications, including object detection, facial recognition, and medical imaging. During training, the model learns by modifying its weights to reduce errors, enabling it to generalize effectively on unseen images.

Numerous retinal conditions, such as diabetic retinopathy, glaucoma, age-related macular degeneration, hypertensive retinopathy, retinal vein occlusion, and retinitis pigmentosa, can be diagnosed via retinal fundus imaging. However, the study focuses on four circumstances in particular: normal retina, tessellation, myopia, and choroidal neovascularization. These specific illnesses are chosen due to their clinical significance and unique visual indicators. Myopia is common and has major effects on retinal health; choroidal neovascularization is linked to severe vision impairment, so early detection is crucial; and tessellation, which is frequently seen in youngsters, is comparatively understudied, indicating a research void. This study aims to create a strong deep learning framework that can accurately classify these particular retinal states by focusing just on these circumstances, which will increase the accuracy of diagnosis.

Particularly with regard to tessellated fundus images, the study fills in a number of important gaps in the existing literature on fundus image classification. Studies on fundus tessellation that have already been done, such as those by Huang et al. (2023) and Shao et al. (2021), tend to concentrate on particular groups, such as children, and report a modest level of accuracy without stressing interpretability. On the other hand, the study model uses data augmentation and interpretability techniques like Grad-CAM and SHAP to reach a high accuracy of 90.8% for tessellated fundus categorization and 97% across four conditions (choroidal neovascularization, myopia, tessellation, and normal retina). Particularly in the classification of tessellated fundus images, this degree of accuracy and model transparency represents a significant breakthrough in the area, addressing a gap in explainability and model performance for larger populations.

Since most studies only use one deep learning model, there is also a noticeable lack of research on ensembling strategies to increase model robustness. The work lays the foundation for future ensembling techniques, which can further improve classification accuracy and resilience in fundus imaging, even though the focus has been on improving a VGG16-based model with class weighting, early halting, and model checkpointing. In addition, there has not been much effort put into converting high-performing models into usable tools for real-world applications in earlier research, which has neglected their practical use. In order to address the requirement for scalable, user-friendly diagnostic tools that can be immediately incorporated into clinical practice, current research tends to produce a mobile or online application in conjunction with a fundus camera that is compatible with both iOS and Android.

Some developments in AI-based ophthalmology include Zedan et al.’s (2023) [30] thorough analysis of deep learning methods for automated glaucoma diagnosis utilizing retinal fundus images, which covered both the main obstacles and the possibilities. In order to increase the robustness of glaucoma diagnosis, Shi et al. (2023) in [31] present an artifact-tolerant contrastive embedding learning system. Ghouali et al. (2022) in [32] investigate teleophthalmology applications that use AI to diagnose diabetic retinopathy, making screening solutions more easily accessible and effective.

In addition, the majority of contemporary research makes use of sophisticated segmentation methods, which can be computationally demanding, such as U-Net or Mask R-CNN. Using a traditional segmentation technique indicates that high accuracy can be attained with less complicated processing requirements, which makes the proposed method more approachable, especially in environments with limited resources. Lastly, even though much research places a high priority on predicting performance, interpretability,which is essential for clinical acceptance, is frequently lacking. The use of Grad-CAM and SHAP to explain model predictions makes this research work more transparent and reliable for clinicians by addressing a significant gap in model interpretability. This study makes a significant contribution to the academic and clinical domains by providing a high-accuracy, interpretable, and valuable solution for fundus image classification.

Step (i). Data Augmentation

Techniques for data augmentation contribute to the diversity of the dataset, which enhances the generalizability of the model.

Rotation (θ) (1)Irotated(x,y)=I(xcos⁡θ−ysin⁡θ,xsin⁡θ+ycos⁡θ)

Flipping

Horizontal Flip: (2)Iflipped(x,y)=I(−x,y)

Vertical Flip: (3)Iflipped(x,y)=I(x,−y)

Rescaling (4)Irescaled(x,y)=I(x,y)max(I)

Step (ii). Splitting the Dataset

Training Set: 80% of the dataset.

Validation Set: 20% of the dataset.

Randomly split data into validation and training sets: (5)Dtrain,Dval=split(D,train_size=0.8)

Step (iii). Transformation of Input Shape

The input images are converted to make them compatible with VGG16 for a consistent (224,224,3) shape : (6)Iinput=resize(I,(224,224,3))

Step (iv). Load the Pre-Trained Base Model: VGG16

Without Top Layers: Remove fully connected layers from VGG16, retaining only the convolutional layers for feature extraction.

Step (v). Freeze the Base Model

To prevent the weights of the VGG16 model’s layers from changing during training, freeze them.: (7)Wfrozen=Wpre-trained(for all convolutional layers)

Step (vi). Add Custom Top Layers

Fully Connected (Dense) Layer: (8)Z=W⋅X+bwhere Z is the output, W are weights, X is input, and b is bias.

Activation Function (ReLU): (9)f(Z)=max(0,Z)

Dropout (with Keep Probability p): Generate Mask: (10)mi∼Bernoulli(p)

Apply Mask: (11)h′=m⊙h

Scale during Training: (12)hscaled′=m⊙hp

Step (vii). Model Compilation

Loss Function (Categorical Cross-Entropy): (13)L=−∑i=1Cyilog⁡(y^i)where C is the number of classes, yi is the true label, and y^i is the predicted probability.

Optimizer (Adam): Parameter Update: (14)θ=θ−αmv+ϵwhere m and v are the first and second moment estimates, α is the learning rate, and ϵ is a small constant to avoid division by zero.

Step (viii). Callbacks

Early Stopping: Keeps track of the validation loss and, after a predetermined number of epochs, stops training if there is no improvement. (15)if ΔLval<ϵ for k epochs, stop training

Learning Rate Reduction on Plateau: If validation loss is not decreased after a predetermined number of epochs, the learning rate is reduced by a factor f. (16)αnew=α×f

Model Checkpointing: Saves the model weights whenever the validation loss achieves a new minimum.

Step (ix). Training the Model

Forward Pass (Feedforward): Calculate activations at each layer. (17)a=f(W⋅X+b)

Backward Pass (Backpropagation): Use the optimizer to update the gradients of the loss with respect to each weight.

Step (x). Testing and Prediction

For a given test image Xtest, the model computes the output probabilities for each class.

Softmax Activation (for Multi-class Prediction): (18)y^i=ezi∑j=1Cezj

Prediction: (19)Predicted Class=arg⁡max(y^i)

Step (xi). Explainability (XAI) Using Grad CAM and SHAP

Grad CAM (Gradiant-Weighted Class Activation Mapping), SHAP (SHapley Additive exPlanations) computes feature importance for each prediction: (20)ϕi=∑S⊆N∖{i}|S|!(|N|−|S|−1)!|N|!(v(S∪{i})−v(S))where ϕi is the Grad CAM & SHAP value for feature i, S is a subset of features, N is the set of all features, and v(S) is the model’s output for subset S.

Step (xii). Save the Model

The trained model’s weights and architecture are saved to facilitate reuse and inference: (21)Save Model→File

Step (xiii). Model Evaluation Metrics

Validation & Loss Graphs: Monitor training and validation set’s accuracy and loss over time.

Classification Report: F1-score, precision, and recall for every class.

Precision: (22)Precision=True PositivesTrue Positives+False Positives

Recall: (23)Recall=True PositivesTrue Positives+False Negatives

F1 Score: (24)F1=2⋅Precision⋅RecallPrecision+Recall

Confusion Matrix: The study represents a confusion Matrix showing true vs. predicted classifications for each class.

Class weights and parameters are employed to modify the loss function to address the class imbalance. (25)Class Weight=Total SamplesNumber of Classes×Samples per Class

Data augmentation approaches were essential in creating a strong model for fundus image classification. By describing how images are rotated by an angle θ. Eq. (1) enables the model to learn from orientation differences, which is crucial for enhancing generalizability by describing how images are rotated by an angle θ. In addition. Furthermore, Eqs. (2) and (3) describe how to flip images both vertically and horizontally, providing a means of adding reflections and thereby artificially expanding the dataset. Images are rescaled using Eq. (4), which normalizes pixel values to fall between [0, 1]. In order to improve the stability of learning during training, this normalization is essential for guaranteeing that the model analyzes images consistently.

As per the system architecture, the flow starts with data acquisition, then preprocessing the images with data augmentation and resizing with different strategies. The model is trained using VGG16 with fine tuning for better performance acquisition, and interpretability using Grad-CAM and SHAP as final predictions for medical experts. The flow diagram Fig. 4 visually conveys the overall architecture and every aspect as some interconnection that interacts within the system. The flowchart gives an in detail the sequential steps of the proposed approach, which will further guide the reader through a pipeline, that highlights the decision points and iterative processes involved in model implementation.

The dataset was then divided into training and validation sets, as shown in Eq. (5) to properly evaluate the model’s performance on unseen data. This divide improves the model’s ability to generalize to new images by reducing overfitting. For consistency in feature extraction, input images must be transformed to a standard shape of (224,224,3) in order to be compatible with the VGG16 model, as described in Eq. (6). The model design is described in depth in Eqs. (7)–(12), which include freezing the VGG16 pre-trained layers to preserve learned features and modified the top layers with dropout layers and dense connections to increase model capacity and avoid overfitting. In particular, Eq. (10) provides dropout regularization, which randomly disables neurons during training to support a more robust learning process, whereas Eq. (8) indicates the functionality of the fully connected layer.

Eqs. (13) and (14) demonstrate how important model compilation is for defining the optimizer and loss function. The model’s predictions are compared to true labels using the categorical cross-entropy loss function to ensure proper categorization. Convergence is accelerated by the Adam optimizer, which is described in Eq. (14) and effectively updates model parameters using first and second-moment estimates. Callbacks such as early stopping and learning rate reduction are emphasized in Eqs. (15) and (16) significantly enhancing training by avoiding overfitting and modifying the learning rate in response to validation performance. Eqs. (18) and (19) show how the softmax activation function is employed to translate logits into probabilities during testing and prediction, allowing the model to produce well-informed class predictions.

For the interpretation of results, Grad CAM and SHAP were utilized to improve the interpretability of the model, as shown in Eq. (20), which measures the contributions of features to the model’s output and provides information about the decision-making process. A thorough examination of classification performance is provided by model evaluation measures such as precision, recall, and F1-score, which are covered in Eqs. (22) through (25). The confusion matrix enables a more nuanced comprehension of the model’s advantages and disadvantages by providing a visual depiction of true vs. anticipated labels. All things considered, each equation makes a substantial contribution to the overall framework for the automatic classification of fundus images, enabling a successful methods of retinal disease detection.

The fundamental parameters and components of the methodology are shown in the Table 2, with particular attention to the crucial factors that influenced the proposed model architecture, data handling, training procedure, and assessment. In order to use a pre-trained VGG16 model as a feature extractor, this study removed its top layers and created bespoke top layers for multi-class classification that targeted images of normal fundus, myopia, tessellation, and choroidal neovascularization. Input images were standardized to 224 × 224 × 3 shapes and data augmentation techniques, such as rotation, shifts, shear, and zoom, were utilized to improve model generalization and solve dataset restrictions.

With a learning rate of 1 × 10−4, the Adam optimizer was used for training, and categorical cross-entropy was utilized to match the multi-class configuration. Class weights were utilized to counteract the imbalance between classes. A number of countermeasures were used, including early stopping and learning rate reduction on plateau to avoid overfitting and improve training effectiveness. Model checkpoints were also saved based on validation loss to maintain ideal weights.

The key hardware and software elements used in the creation and use of a fundus image classification model for the diagnosis of four distinct retinal conditions–choroidal neovascularization, myopia, tessellation, and normal–are compiled in Table 3. The Mendeley-sourced dataset included the annotated photos required for model evaluation and training. In addition to local computing resources for preliminary development and testing, Google Colab’s GPU support was crucial for faster training. In addition, a fundus camera is intended to collect data in real-time, enabling ongoing model validation under actual circumstances.

The project mostly used Python 3.9.0, which was controlled by Miniconda and made use of the TensorFlow and Keras libraries to construct and refine a VGG16-based model that was trained on ImageNet. Essential tools such as class weight handling and metrics computations were handled by Scikit-Learn, while data visualization and model performance analysis were made easier with Matplotlib and Seaborn.

Despite advanced models such as Resnet, Densenet and Efficientnet that provide improved performance by increasing complexity and computational demands. Due to the shortage of lower dataset which has limited images, the study can encounter deeper architecture risk overfitting, whereas VGG16 which has a proper regularization such as dropout layers, early stopping and data augmentation mechanisms, will achieve strong generalization in classification.

The major concern about using VGG16 was based on its proven effectiveness in clinical and medical image processing mechanisms towards its simplicity and robustness. Since it is an older architecture that demonstrated strong feature extraction methods in fundus images for classification. In addition it also exhibits structured and layer wise feature representation, that provides models interpretability using explainable artificial intelligence (XAI).

Grad-CAM and SHAP were utilized to improve interpretability for each classed condition by offering visual insights into model predictions. Flask and Ngrok will be used in future deployments as web or mobile applications, enabling real-time diagnostic capabilities and increased accessibility. This thorough setup guarantees that the project has all the tools required for both reliable model creation and possible real-world application.

A collection of annotated retinal fundus photos was obtained from Mendeley [41], is depicted in Table 4, and it is the source of the dataset used in this investigation. These pictures include a wide range of classes, including Normal, Myopia, Choroidal Neovascularization, and Tessellation, which makes it easier to train a thorough model for automated disease classification.

This study employed a dynamic class weighting scheme to address the significant class imbalance in the dataset. This targeted strategy for balancing class representation is critical for medical applications where underrepresented conditions must be reliably detected.

The distribution of training and validation fundus images among the dataset’s various classes is shown in the Table 5. It shows how many photos are allotted to each class, such as normal, tessellated, myopia, and choroidal neovascularization.

The percentage of training dataset, validation dataset and testing dataset of fundus images is shown graphically in the Fig. 5.

Since the dataset size is less, the study consideres augmentation techniques which include, rotation, flipping, contrast adjustments, resizing and zoom transformation. These techniques expanded the dataset by effectively raising the variability and reducing data imbalance. The study compared toout augmentation to exhibit the impact of augmentation. It obtained an accuracy of 72.3% without augmentation. In addition to that, validation accuracy was raised to 77.5% from 95.3% with F1-score of 0.50% from 0.36%.

The study focuses on the evaluation using the base model VGG16, which is a convolutional neural network architecture with 16 layers–13 convolutional layers and 3 fully connected layers–that is renowned for its deep structure and simplicity. It is useful for feature extraction in image classification applications since it uses small 3 × 3 filters and has a standard architecture. VGG16 is a reliable base model that has been pre-trained on the ImageNet dataset. It is frequently adjusted for particular uses such as fundus picture classification and medical image analysis.

The high-level features that the network learned from the input photos of various retinal states are visualized by the feature maps that were extracted from the VGG16 model is shown in Fig. 6. Specific patterns, textures, and forms that the model determines are important for classification are captured in each feature map. These feature maps, for example, can show how the model differentiates between minute changes in retinal structure and anomalies in the context of fundus pictures that depict choroidal neovascularization, myopia, normal conditions, and tessellated patterns. The feature maps’ varied patterns demonstrate how the model can concentrate on important details such as vascular alterations, pigmentation variances, and structural anomalies specific to each ailment.

Examining these characteristics improves interpretability and helps comprehend the model’s decision-making process, both of which are essential for clinical applications in ophthalmology. All things considered, the feature maps that are displayed demonstrate how well the VGG16 design captures the fine features necessary for precise retinal disease classification.

The results vary from the model alone and revised parameters upgradation. Even after the successful upgradation of the model, there is no correct prediction and unsatisfactory results are formulated. This study uses fine-tuning in the last Test case, which has a satisfactory results and appropriate predictions are shown.

In this study of fundus image classification, several techniques were employed to manage and mitigate overfitting: 1.

Data Augmentation: Using techniques such as rotation, flipping, and rescaling,increases the effective size and diversity of the dataset, which helps the model generalize better to unseen data.

The current model’s shortcomings are mostly caused by the small dataset size (302 photos in four classes), which raises the possibility of overfitting and restricts the model’s capacity to generalize to larger populations. In addition, the model can still be skewed toward more frequent classes even when class weighting is used. The use of a single VGG16 architecture can limit the investigation of better models or ensembling methods that can enhance performance. The full diversity of real-world fundus images can not be captured by simple image augmentation techniques, and although Grad-CAM provides some interpretability, more sophisticated methods are required to improve comprehension, especially for medical applications.

The generalizability of the model will be improved for future iterations by adding more varied and annotated fundus photos to the dataset. Using ensembling techniques and experimenting with sophisticated designs such as ResNet, DenseNet, or EfficientNet can increase classification accuracy. Real-world changes can be more accurately simulated by using more sophisticated picture augmentation techniques and creating synthetic data. Transparency will be increased by addressing class imbalance with synthetic data and incorporating sophisticated interpretability techniques such as SHAP or LIME. The model will be improved and made practically applicable in diagnostic situations by real-world testing, mobile app deployment, and clinical validation in conjunction with ophthalmologists.

This study also aims to explore model ensembling techniques to further enhance performance and reliability, since it acknowledges its potential benefits for improving the robust performance in multi-class classification. However given the dataset constraints and main target on interpretability of architecture for clinical applications. The main priority is for the well-optimized model.