The dataset considered in this study consists of retinal fundus images, which were pre-processed with an isotropic Gaussian filter to remove noise and improve clarity. The original dataset was obtained during the APTOS 2019 Blindness Detection challenge (17). Each image was labeled with the severity of diabetic retinopathy (DR) via the train.csv file that came along with the original dataset. Images are arranged into five separate folders, each corresponding to one DR stage. The class-wise distribution is given in the Table 1.

The Diabetic Retinopathy (DDR) dataset comprises 13,673 retinal fundus images collected from 147 hospitals across 23 provinces in China (18). Each image is assigned one of five classes depicting diabetic retinopathy (DR) severity: No_DR, Mild, Moderate, Severe, and Proliferative_DR. The original dataset included a sixth category corresponding to poor-quality images, which were excluded in this study to ensure reliable training. This brings the total number to 12,522 images. Also, all of the images were preprocessed to remove the black background since it boosted the visibility of retinal structures and would have taken away from irrelevant information. The distribution of images in the five DR severity levels is given in Table 2. Figure 2 shows fundus images from the dataset.

Potential confounding factors such as patient age, gender, geographic region, and imaging device variability were addressed at the dataset acquisition level. The DDR dataset was collected from 147 hospitals across 23 provinces using standardized acquisition principles and diverse fundus camera types, with balanced demographic representation. This multicenter and multi-device design reduces systematic bias and improves generalizability. In both the DDR and APTOS 2019 datasets, diabetic retinopathy severity labels serve as the primary stratification variable, reflecting clinically meaningful disease progression. As the study is image-centered and does not involve patient-level intervention or outcome modeling, statistical confounding adjustment methods such as propensity score matching are not directly applicable.

The sample size used in this study was determined by the availability of publicly accessible benchmark datasets, namely APTOS-2019 and DDR, which are widely adopted in diabetic retinopathy research. As this work follows a data-driven deep learning framework rather than a hypothesis-testing statistical design, conventional sample size calculation, effect size estimation, and power analysis are not directly applicable. Model validity is instead assessed through generalization performance on independent test sets and consistent comparative evaluation across multiple architectures and datasets. This approach is standard practice in contemporary medical image analysis research.

CheXNet refers to a deep CNN model designed specifically for the automated identification of pneumonia and thoracic illnesses from chest X-ray images (19). The model is based on DenseNet121, a 121-layer deep architecture, which introduces dense connections between all the layers and their preceding layers. This direct linkage of layers enables a layer to access the feature maps of all previous layers at once, which improves the gradient flow, promotes reuse of features and has fewer parameters than a conventional deep CNN of the same depth. This architecture helps a lot with medical imaging since it recognizes very subtle patterns at multiple scales. DenseNet121 is a construction of dense blocks with transition layers, such that in a dense block, each layer takes the concatenation of the feature maps of all preceding layers as input. Formally, for a layer l with input feature maps x₀, x₁, …, x_l−1, the output x_l is:

Where ([x₀, x₁, …, x_l−1]) is the concatenation operator and H_l(·) is a composite function of Batch Normalization, ReLU activation, and a 3 × 3 convolution. Thus, this setup allows the network to blend low- and high-level features efficiently to pick up subtle patterns in medical images. Transition layers lie between dense blocks to shrink the size of feature maps and reduce channel numbers to increase computational efficiency and counter overfitting. Each transition layer carries out 1 × 1 convolution with 2 × 2 average pooling as follows:

After the final dense block, CheXNet applies Global Average Pooling (GAP) to reduce each feature map to a single value (20). The resulting vector is then passed through a fully connected layer with sigmoid activation for multi-label classification:

where C is the number of disease classes, wiTand b_i are learnable weights and biases, respectively, and σ(·) gives the probability of each class between 0 and 1. CheXNet is trained by means of binary cross-entropy loss so that the network can output probabilities closely reflecting actual presence and absence of each disease. There are a number of limitations of the CheXNet though. It belongs to the class of domain-specific algorithms because it has been originally trained on chest X-ray images; thus, direct application to other modalities, such as retinal images, is suboptimal. DenseNet (21), on the other hand, mainly extracts local features and thus it cannot maximally utilize the long-range spatial relation between features. Either way, CheXNet, with its dense connectivity, multi-scale feature reuse, and deep design, provides a very good backbone for analysis of medical images. Nonetheless, applying it to a new domain usually requires modifications to the architecture in terms of attention modules, stronger regularization, and improved interpretability techniques. In short, to make CheXNet work for retinal fundus images, it must be channeled to outweigh certain clinically relevant eye regions and to capture long-range dependencies. Hence, attention modules, hybrid CNN-transformer architectures, and other specialized training techniques are some of the improvements that must be considered for reliable classification of eye diseases. These improvements and architectural changes are described in detail in the next section. Figure 3 shows the hierarchical structure of the CheXNet model.

Although CheXNet was originally proposed for chest X-ray image analysis, in this work, it is adapted to the retinal fundus image domain through transfer learning and architectural modification. The DenseNet121 backbone of CheXNet is initialized using pretrained weights, while the original classification layer—designed for thoracic disease detection—is removed and replaced with a task-specific fully connected layer tailored to diabetic retinopathy classification. The network is then fine-tuned on retinal fundus datasets, allowing the learned low-level and mid-level visual representations to be adjusted to the distinct texture, color distribution, and lesion patterns of fundus images. Furthermore, to mitigate the domain gap between chest X-ray and retinal imaging modalities, a Convolutional Block Attention Module (CBAM) is incorporated into the architecture. This enables the model to emphasize clinically relevant retinal structures such as microaneurysms, hemorrhages, and exudates, while suppressing irrelevant background information. Through this adaptation strategy, CheXNet serves as a robust and transferable backbone rather than a domain-restricted model, making it suitable for accurate diabetic retinopathy diagnosis.

Unlike existing diabetic retinopathy studies that primarily apply attention mechanisms directly to generic DenseNet-based architectures or design task-specific CNNs from scratch, this work introduces a novel adaptation of the CheXNet architecture—originally developed and validated for chest X-ray analysis—into the retinal fundus imaging domain. The novelty of the proposed CheXNet_CBAM model lies in three key aspects. First, it repurposes CheXNet as a medical-image backbone beyond thoracic imaging and systematically enhances it with a convolutional block attention module (CBAM) to address the unique spatial and pathological characteristics of retinal fundus images. Second, the integration of CBAM enables joint channel-wise and spatial attention within a densely connected architecture, allowing the model to emphasize clinically relevant retinal regions (e.g., microaneurysms, hemorrhages, and exudates) without introducing additional supervision or lesion annotations. Third, the proposed framework is extensively validated across two heterogeneous public datasets (APTOS-2019 and DDR), demonstrating strong cross-dataset generalization, which is rarely examined in prior attention-augmented DenseNet or CheXNet-based DR studies. Furthermore, the combined use of CBAM with Grad-CAM visualization enhances model interpretability, providing clinically meaningful explanations that support practical deployment. Together, these contributions distinguish the proposed CheXNet_CBAM model from existing attention-based DR approaches and highlight its incremental methodological and clinical value.

In the proposed model, we highlight a number of modifications that have to do with CheXNet in such a way that it becomes a suitable tool for retinal fundus image analysis. As a first step, various data preprocessing and augmentation techniques have been performed to alleviate the problems caused by the variability of images and the imbalance of classes. After that, a Convolutional Block Attention Module (CBAM) is added to the model to inform the network of clinically significant retinal areas. To conclude, DropBlock regularization is introduced as a means of preventing overfitting and improving the model's capability of generalization. Additionally, Transformer encoder layers are added after the feature extraction from convolutional networks to allow the model to capture long-range dependencies and global structural patterns. The learning rate, dropout, and block size are among the hyperparameters whose careful tuning ensures the training stability. Lastly, for the sake of model interpretability, enhanced Grad-CAM visualization is employed to delineate regions that are specific to the lesion, and this, in turn, provides clinical transparency. In total, these changes raise the baseline CheXNet applied to retinal disease classification in terms of accuracy, robustness, and reliability to a considerable extent.

In the grading of diabetic retinopathy (DR), images generally fall into one of five severity categories: no DR, Mild DR, Moderate DR, Severe DR, and Proliferative DR. No DR here indicates that the retina appears normal and there are no pathological changes, whereas Mild DR indicates early-stage changes such as few microaneurysms or just minor hemorrhages. Moderate DR indicates more widespread pathological changes such as multiple hemorrhages or small exudates. Severe DR indicates considerable retinal damage with diffuse hemorrhages and, of course, vascular lesions; Proliferative DR describes the final stage, which is marked by abnormal neovascularization that can potentially lead to the loss of vision.

Evaluation metrics provide guarantees on the efficiency with which the model can perform on the test classes. Accuracy refers to the ratio of correctly classified images over the total number of images. Precision for a class tells us how many of the predicted cases for that class were correct. Sensitivity (Recall) measures the ability of the model to detect all true cases of a class, while specificity measures how well the model can identify that class from others. F1-score provides a trade-off between precision and sensitivity and is specifically important for the classes that are least common, like Proliferative DR.

For instance, a model may have a high positive predictive value but a low sensitivity for Proliferative DR, meaning that when it predicts this class, it is usually correct but will miss many true cases. Conversely, Mild DR is often more difficult to detect; hence, the importance of the F1-score to verify that mild cases are indeed detected, while minimizing false positives. The relationship between the metrics and each DR severity class is summarized in the Table 3 below.

Finally, 95% confidence intervals for accuracy were estimated using the standard error of the proportion to quantify the statistical reliability of results:

All the trials were conducted with the help of Visual Studio Code (VS Code) on Windows 11 Pro. The system on which training and evaluation were performed had an Intel Core i7-12700K processor, 16GB RAM, and NVIDIA RTX4060ti GPU.

Table 4 presents the performance of the proposed CheXNet_CBAM model compared to various baseline models when applied to fundus images from the APTOS 2019 dataset for diabetic retinopathy (DR). The CheXNet_CBAM model achieved an accuracy of 96.12%, a precision of 96.30%, and an F1 score of 96.08%, outperforming the other models. The CheXNet model without CBAM achieved an accuracy of 93.80%, demonstrating the effectiveness of the proposed model and the idea of adding a CBAM layer to the CheXNet model, which significantly increased the performance of the CheXNet_CBAM model in diabetic retinopathy detection. Meanwhile, the MobileNetV2 model ranked second in diabetic retinopathy detection with an accuracy of 95.02% and an F1 score of 95.01%. The DenseNet121 model also achieved an accuracy of 94.24%. While VGG19 and ResNet50 performed the worst among the models in their ability to detect diabetic retinopathy, with VGG19 achieving an accuracy of 69.99%, ResNet50 ranked last with an accuracy of 52.71%. Figure 5 illustrates the model's effectiveness.

The confusion matrices in Figure 6 demonstrate the superior diagnostic accuracy of the proposed CheXNet_CBAM model compared to other models, achieving high diagnostic accuracy across all five disease categories (mild, moderate, non-diabetic retinopathy, proliferative, and severe). The model demonstrated exceptional ability to differentiate diabetic retinopathy cases, with significantly lower false positives and false negatives compared to CheXNet, DenseNet121, MobileNetV2, VGG19, and ResNet50. This significant performance improvement is attributable to the integration of the CBAM attention mechanism (convolutional block attention module) into the CheXNet architecture, which enables the model to more accurately focus on critical regions in retinal images containing signs of diabetic retinopathy.

The evaluation results of the proposed CheXNet_CBAM model by category, shown in Table 5, demonstrated exceptional performance in diagnosing all stages of diabetic retinopathy. It achieved a perfect precision of 100% in identifying cases (No_DR) with a high recall rate of 97.24%, demonstrating its excellent ability to avoid misdiagnosis of benign cases. The model also demonstrated excellent balance in diagnosing different disease stages, achieving high F1 rates ranging from 92.04% for moderate cases to 97.51% for proliferative cases. It also achieved outstanding performance in diagnosing severe cases with a perfect recall rate of 100% and a precision of 94.24%. These results confirm the model's high ability to accurately distinguish between different disease stages without missing any disease (especially severe and mild), while minimizing misdiagnosis rates.

Table 6 and Figure 7 provides a comprehensive overview of the performance of the proposed CheXNet_CBAM model compared to other baseline models when analyzing fundus images for diabetic retinopathy detection using the DDR dataset. The proposed model outperformed other models in detecting diabetic retinopathy with an accuracy rate of 96.33%, a precision rate of 96.29%, and an F1 score of 96.30%. The CheXNet model without CBAM achieved an accuracy of 91.35% and a precision of 91.18%. MobileNetV2 ranked second with an accuracy of 91.51% for detecting diabetic retinopathy. DenseNet121 and VGG19 achieved an accuracy of 84.42 and 64.06%, respectively. ResNet50 performed the worst among the models in detecting diabetic retinopathy with an accuracy of 41.72%.

ResNet50 performance was low relatively with the APTOS-2019 and DDR datasets. Though the ResNet-based models have been consistently strong in diabetic retinopathy studies, their performance under hyperparameter tuning, training duration, and dataset-specific optimization is very dependent. For conducting our experiments, we had all baseline models under same conditions to maintain a fair and unbiased comparison. One reason for the poor performance of ResNet50 is its restricted use of features and lack of pronounced attention mechanisms, both of which are vital for picking up small and scattered retinal lesions. DenseNet-based architecture, on the contrary, even more so when enhanced with CBAM, permits a richer aggregation of multi-scale features and greater attention paid to the clinically significant areas. The factors combined, besides the proposed CheXNet_CBAM model's robustness and generalization capability across different fundus image datasets, have made it possible to further emphasize performance gap.

Exceptional performance was achieved by the proposed CheXNet_CBAM model, which was presented in Figure 8 and unambiguously surpassed all other models in diagnosing diabetic retinopathy using the DDR dataset. The confusion matrix showed a perfect distribution of values for each class with the highest accuracy for the correct classification achieved. The main diagonal of the confusion matrix had the following values: 626 for severe cases, 634 for proliferative cases, 583 for non-proliferative cases, 558 for moderate cases, and 628 for mild cases. The CheXNet model made more misclassifications than the other models and those misclassifications were especially between adjacent classes. MobileNet and DenseNet121 had a relatively good performance, but their classifications overlapped to some extent. VGG19 and ResNet50, on the contrary, performed very poorly with misclassifications being widely spread across the classes. This clearly indicates the superior performance of the proposed CheXNet_CBAM model, which integrates the CBAM attention mechanism in overcoming high diagnostic accuracy and low error rates.

The CheXNet_CBAM model demonstrated exceptional ability to detect both mild and severe cases of diabetic retinopathy, achieving a perfect recall rate of 100% for both categories, as shown in Table 7. In severe cases requiring urgent medical intervention, it achieved an extremely high precision of 98.74%, detecting all cases, while in mild cases requiring early follow-up, its precision reached 97.21%. Severe proliferative cases were treated with exceptional precision, achieving an accuracy rate exceeding 99.36% across all metrics. Even in more complex cases, such as moderate and benign cases, the model maintained a strong performance exceeding 90%, confirming its ability to distinguish between all stages of the disease accurately. This balanced performance makes it a reliable medical tool that clinicians can rely on to make accurate and informed treatment decisions.

The statistical analysis in Table 8 reveals a clear performance hierarchy, with the CheXNet_CBAM model outperforming with an accuracy of 96.12%. CheXNet_CBAM shows a statistically significant improvement of 2.32 percentage points due to the integration of CBAM, compared to the CheXNet model with non-overlapping confidence intervals (95.23, 96.98) vs. (92.75, 94.82). The 1.10 percentage point difference represents a significant but smaller gap with the MobileNetV2 model, with slightly overlapping confidence intervals indicating moderate statistical significance. The huge performance gaps 26.13% vs. VGG19, 43.41% vs. ResNet50 show highly significant differences with non-overlapping confidence intervals.

Improved Performance Gaps: the DDR dataset shows more pronounced performance differences, as shown in Table 9, with CheXNet_CBAM achieving 96.33% accuracy and larger gaps compared to the other models. The 4.98 percentage point improvement between CheXNet_CBAM and CheXNet is highly statistically significant with completely non-overlapping confidence intervals of (95.46, 97.18) vs. (90.18, 92.50). The 4.82 percentage difference between CheXNet_CBAM and MobileNetV2 demonstrates clear statistical significance. The 11.91 percentage point difference is statistically significant with non-overlapping confidence intervals. The 11.91 percentage point difference demonstrates statistical significance with non-overlapping confidence intervals between CheXNet_CBAM and DenseNet121.

The statistical analysis results provide strong evidence of the superiority of the proposed CheXNet_CBAM model, achieving high performance on both datasets and demonstrating significant differences from all competing models. Clinical reliability, thanks to precise confidence intervals, ensures predictable performance, with the model's generalizability. Overall, the inclusion of confidence intervals provides clinically meaningful effect size interpretation and supports the reliability and generalizability of the proposed CheXNet_CBAM framework.

The meaning of the trend is the development of the classification of diabetic retinopathy that is prevalent in its various features, from methods, from feature fusion to attention-based techniques. For instance, RT2Net integrated fundus and vascular branch networks (41) and achieved 88.2% on EyePACS and 85.4% on APTOS, while few-shot learning with Siamese neural networks and pre-trained models such as VGG16 and ResNet50 (42) reported more modest results (80%−81%).

Such results were shown by EfficientNetB3-enhanced models with SE processing blocks (43), which achieved an 88.44% accuracy rate. Attention-driven methods, such as MSCAS-Net (44), have also greatly raised this performance level to an astounding 93.8% on the APTOS database, surely proving how important multi-scale and attention fusion are in performance improvement. Therefore, it is evident that the introduction of attention mechanisms and feature-level fusion is more advantageous when compared with standard transfer learning techniques. Other studies have concentrated on interpretability and hybrid learning for the sake of accuracy vs. clinical applicability. With an amazing 94.64% on APTOS, the explanation AI is from ResNet-50 with SHAP (45) compared with traditional CNN attention frameworks such as MSRAB + CrAB (46), which achieved 88.31% on APTOS. Bayesian deep learning approaches (47) achieved the highest performance, with 94.23% accuracy using MC Dropout, emphasizing their strength in uncertainty-aware classification. Hybrid CNN–ViT models with interpretability tools such as LIME and Grad-CAM (48) also demonstrated high performance (93.01%), showing that combining convolutional and transformer-based features yields both strong results and interpretability.

In the meantime, hybrid strategies that fuse deep features with classical machine learning did not do that well, compared to the few, but for example, an SVM classifier combined with EfficientNetV2-S (49) worked at 91%, while the lightweight two-stage models (50) had an accuracy of 90.75%-which is just a marginal improvement. Swin-TransformerV2 with hybrid attention (51) showed performance with average accuracy (85.5%−87.9%), not so bad compared to CNN-driven models. Nonetheless, our proposed model, CheXNet_CBAM, which incorporates methods like DenseNet121 with CBAM, DropBlock2D, and a Transformer Encoder, attained 96.12% accuracy on APTOS and 96.33% accuracy on DDR-with the rest of the models competing with the Bayesian best ones.

The most salient points of improvement to CheXNet include the focus on clinically relevant retinal regions that was made possible with the incorporation of the convolutional block attention module (CBAM) within the model, DropBlock2D for greater regularization control and preventing any tendencies to overfit, and the use of a Transformer Encoder to capture contextual and long-range spatial dependencies in retinal images. Such a hybrid model stands to combine the strengths of contextualization with CNN-based feature extraction via attention mechanisms and transformers. Therefore, CheXNet_CBAM is not only better in terms of accuracy than most state-of-the-art methods, but also provides a sound and generalizable framework, making it a prime candidate for real clinical deployment. Table 10 shows a comparison of previous studies on the diagnosis of diabetic retinopathy.

Recent multicenter studies have demonstrated the effectiveness of deep convolutional and attention-based models for diabetic retinopathy screening across heterogeneous populations and imaging devices. While these works primarily focus on large-scale performance benchmarking, limited attention has been given to integrating lightweight attention mechanisms with dense feature reuse and explicit visual interpretability. In contrast, the proposed CheXNet_CBAM framework emphasizes both classification performance and clinical transparency through Grad-CAM visualization, while maintaining robustness across multiple public datasets. This complementary focus distinguishes the present study and contributes incremental value to existing multicenter research.

Figure 9 presents Grad-CAM visualizations of the proposed CheXNet_CBAM model when applied to fundus images from the DDR and APTOS 2019 datasets, respectively. These visualizations provide insights into the regions of interest emphasized by the model during diabetic retinopathy (DR) detection, highlighting how the network interprets subtle to severe pathological changes.

In Figure 9A (APTOS 2019 dataset), the Grad-CAM output is compared to DDR and displays more pronounced and localized activations, which is probably due to the better image quality of the dataset. The network highlights the small changes around the macula in case of mild DR, but in the case of moderate DR, it reveals more accurately the presence of small hemorrhages and vascular malformations. In No_DR, the distributions of activations stay at the same low levels and are dispersed, which is an indication of a correct diagnosis. In proliferative DR and severe DR, the visualizations highlight pathological areas, including abnormal vascular growth and widespread hemorrhages, confirming the model's ability to identify important clinical features in advanced cases.

In the DDR dataset shown in Figure 9B, the Grad-CAM maps indicate that in case of Mild DR the model's attention is mostly given to the macular area where early retinal changes are taking place. During Moderate DR, the focus is progressively broader and includes microaneurysms and hemorrhages which are in places across retina. No_DR condition is characterized by faint and weak activations, in line with the lack of pathological conditions. On the contrary, the network highlights the areas of neovascularization very well for Proliferative DR, while for Severe DR it points to the areas of extensive bleeding and exudates, giving out intense activations all over the infected zones. Generally, these Grad-CAM results indicate that the suggested CheXNet_CBAM model not only provides better accuracy in classification but also gives clinically significant interpretation. The differences between DDR and APTOS 2019 datasets reveal the impact of dataset quality on the not so clear model visualization for DDR and the very clear attention maps for APTOS. The interpretability that comes with this model increase its acceptance in the field of eye care since it allows ophthalmologists to follow and even trust the network's decision-making process.

The Grad-CAM visualizations consistently highlight anatomically and clinically meaningful regions, including areas around the macula, optic disc, and vascular structures where diabetic retinopathy lesions are known to occur. These activation patterns align with established ophthalmic diagnostic criteria reported in the literature, supporting the clinical plausibility of the learned representations. Although quantitative expert annotation-based validation was not available for the utilized public datasets, the observed consistency across datasets and severity levels provides indirect evidence of interpretability reliability.

The proposed CheXNet_CBAM framework is intended as a clinical decision-support system (CDSS) for diabetic retinopathy (DR) screening/triage, rather than as an autonomous diagnostic tool. In a hospital setting, the workflow would be integrated into routine ophthalmic imaging processes. Retinal fundus images would be acquired using non-mydriatic or mydriatic fundus cameras according to local clinical protocols and transferred to an on-premises, access-controlled inference server over the hospital intranet in standard formats (e.g., JPEG or PNG). Prior to inference, images would undergo automated preprocessing consistent with the training pipeline, including resizing to the model input resolution, applying the predefined normalization procedure, and applying an image-quality control step to identify ungradable or non-conforming images (with explicit criteria to be reported). Inference would be executed within a version-controlled software environment (e.g., Python with TensorFlow or PyTorch) using pinned library versions to improve reproducibility across deployments. The system would output both the predicted DR grade and Grad-CAM heatmaps highlighting image regions that most strongly influence the model output, and these outputs would be presented to clinicians via a graphical user interface to support review and documentation. The predictions and justification would be reviewed by the clinicians and they would be able to confirm, override or ask for an image to be re-taken depending on the local practice; the clinical responsibility stays with the physician who finally decides about the patient's management. A structured onboarding would be provided to the users covering the intended use, the interpretation of outputs, and the limitations of AI-assisted screening. Quality assurance would include performance monitoring at periodic intervals using predefined validation cases and acceptance criteria, logging of predictions and model versions for auditability, and targeted review of attribution maps to flag implausible patterns while noting that saliency methods do not establish causal relevance. The appropriate response to suspected performance degradation (e.g., due to acquisition changes) would include manual review and, if justified, a governed model update process with re-validation prior to redeployment. Although evaluation through public datasets can demonstrate generalization in controlled conditions, device- and site-specific external validation is still required to ensure safe integration into heterogeneous real-world screening workflows.

Regardless of the stated effectiveness, the CheXNet_CBAM model that has been suggested has limitations that the researchers consider very important. To begin with, the model was trained and tested on datasets that are publicly available which could be a reason for not reflecting completely the heterogeneity and case-mix of real-world patient populations met in the clinics. The differences in age distribution, disease prevalence in different areas, and other characteristics of the population may lead to the model being less accurate and less reliable when used in hospital settings different from those of the assessed datasets. Second, the assessment was retrospective and based on pre-existing datasets; no prospective workflow integration or evaluation of decision impact was performed. Consequently, downstream clinical outcomes—such as the appropriateness and timeliness of referral pathways—could not be assessed. Prospective, site-specific external validation in real clinical environments is therefore necessary to characterize clinical utility and safety. Third, as with AI-assisted screening systems generally, potential safety risks remain when translating the approach into practice, including misclassification of disease severity and inappropriate triage of higher-risk cases. The proposed method is intended to function as a decision-support system rather than an autonomous diagnostic tool, and clinical responsibility remains with qualified clinicians who make final management decisions. Although Grad-CAM heatmaps can provide qualitative attribution cues that may support clinical review, saliency-based explanations do not guarantee faithfulness and should not be treated as a standalone indicator of prediction reliability. Finally, deployment depends on operational infrastructure and governance, including secure integration into clinical workflows, technical support, fallback procedures for manual screening, and ongoing user training. Because data-driven models may also reflect dataset and labeling biases, ongoing monitoring, periodic validation, and recalibration as needed are required to maintain performance across evolving imaging devices, acquisition protocols, and clinical practices. Collectively, these considerations may hinder translation unless addressed through rigorous external validation and deployment governance.