Integrated Deep Learning Framework for Retinal Disease Detection Using Advanced CNN and Imaging Analysis

Yang et al. [16] intended pre-trained CNN architectures—VGG16, DenseNet121, ResNet50 and ResNet152—for COVID-19 identification utilizing X-ray and CT-scan images. In CT-based binary classification, Fast. AI ResNet autonomously adjusted architecture, preprocessing and training to achieve over 96% accuracy and F1 score. In X-rays, enhanced VGG16 distinguished COVID-19 and pneumonia with 99% accuracy.

Rana and Bhushan [17] evaluate medical imaging-based ML and DL algorithms for illness identification and categorization. Experimental validation was performed on 40 primary studies using MRI datasets. DL models outperformed ML classifiers and CNN architectures in large medical datasets in terms of accuracy and scalability.

Wang et al. [18] offered a bibliometric study examining 2,685 original research articles from 2012 to 2020 to measure trends in medical image analysis DL. U-Net, ResNet, VGG and GANs were employed to segment and classify thoracic, abdominal and retinal scan datasets. Since 2017, DL applications have increased, with transfer learning dominating studies related to COVID-19.

Li et al. [19] examined medical image analysis using CNNs, RNNs, LSTMs, GANs and hybrid models. The study assesses algorithms on publicly available MRI, CT and X-ray datasets. CNNs are the most used approach, and Python is the chosen programming language. The accuracy and resilience of classification and segmentation tasks were often high.

The improved deep retinal intelligence detector network (IDRiD-Net) detects 16 retinal conditions from optical coherence tomography (OCT) images using a DL algorithm with uncertainty estimates, trained on massive, multi-device datasets, including internal and external test sets, as analyzed by Peng et al. [20]. It achieves an F1 score of up to 97.44% with a thresholding strategy and flags over 85% of nontarget or low-quality photos for manual review, outperforming state-of-the-art methods and human specialists.

Patel et al. [21] classified lung X-ray images using EfficientNet-B4 transfer learning and Grad-CAM for explainable multi-disease classification. We trained the model on a dataset of lung X-rays featuring COVID-19, pneumonia, tuberculosis and pulmonary nodules. Grad-CAM visuals improved interpretability and classification accuracy to 96%.

Sarah et al. [22] classified OCT images into diabetic macular edema, drusen and choroidal neovascularization using CNNs with various optimizers and regularization methods. The collection includes over 80,000 annotated retinal scans from Kaggle's OCT picture dataset. The model achieved a classification accuracy of over 98%, with high precision and recall.

Tayal et al. [23] introduced an OCT retinal scan–based DL diagnostic system utilizing CNNs with five, seven and nine layers to classify four ocular conditions: diabetic macular edema, drusen, choroidal neovascularization and normal retina. OCT images preprocessed for noise removal, contrast enhancement and contour identification are in the collection. The proposed models outperformed manual diagnosis in classification accuracy (96.5%), sensitivity (96.0%) and specificity (98.6%).

Diwan et al. [24] examined the You Only Look Once (YOLO) family of object detection models, which use regression-based formulations to recognize objects in a single network run. The study evaluated several YOLO variants on the COCO and PASCAL VOC datasets. The results showed that YOLO achieved up to 63.4% detection accuracy while providing inference speeds approximately 300 times faster than two-stage detectors such as Fast R-CNN, which achieved around 70% detection accuracy.

Valanarasu et al. [25] implemented KiU-Net, a dual-branch segmentation architecture, which uses an overcomplete network (Kite-Net) for fine details and U-Net for high-level feature learning. The suggested technique achieved strong segmentation with fewer parameters and faster convergence on five varied medical imaging datasets. The results revealed greater boundary precision and small structure recognition than typical U-Net variations. The KiU-Net 3D extension improved volumetric segmentation.

Despite significant progress in DL-based retinal disease detection, existing studies primarily focus on isolated CNN architectures, limited datasets and single-disease classification without addressing clinical interpretability and deployment challenges [26,27,28,29]. Most methods lack robust generalization across diverse imaging conditions, do not integrate explainable AI tools for clinician trust and are rarely designed with teleophthalmology support for use in remote healthcare environments. These gaps restrict scalability and limit their clinical applicability.

Unlike existing studies that focus on single-disease classification, limited datasets or non-explainable CNN models, this research introduces a unified and clinically oriented framework that combines optimized transfer learning, multilevel retinal disease grading and Grad-CAM–based interpretability. The proposed TL-Pretr-CNN model is designed to generalize effectively across heterogeneous high-resolution fundus images while providing transparent visual explanations to support clinical decision-making. This integrated approach, along with teleophthalmology compatibility, establishes a stronger practical contribution compared to prior literature and demonstrates the system's applicability in real-world healthcare environments.

The study utilizes a large-scale retinal fundus image dataset containing 130,993 images, each annotated with one of five DR severity levels: No DR, mild, moderate, severe and proliferative DR. This dataset is sourced from a publicly available benchmark repository widely used in DR classification research. The size discrepancy mentioned earlier was due to partial subsets being described during the preprocessing explanation; however, the complete dataset used for training, validation and testing consists of 130,993 images, which is consistently applied throughout the model development pipeline.

All images were captured using high-resolution fundus cameras under varying illumination and field-of-view conditions, reflecting real-world clinical variability. The dataset offers a balanced representation across DR stages, though mild class imbalance natu-rally exists in higher severity grades—typical of epidemiological DR distributions.

To mitigate the class imbalance, stratified sampling was employed to split the dataset into training (70%), validation (15%) and testing (15%) sets, ensuring proportional representation of all five DR classes in each subset. This preserves clinical representativeness and prevents bias during model optimization. In addition, data augmentation was applied preferentially to underrepresented classes, including contrast-limited adaptive histogram equalization (CLAHE)-based contrast enhancement, rotations (±20°), random cropping and brightness/contrast adjustments, to artificially balance the training data and improve generalization. While explicit k-fold cross-validation was not performed, the separate validation set was used for hyperparameter tuning and monitoring overfitting, and the final testing set provided an unbiased evaluation of model performance.

Each image, provided in RGB format with varying dimensions, was resized to a fixed resolution as part of preprocessing. Normalization, augmentation and contrast enhancement via CLAHE were applied as follows: Iinorm=Ii−μσ I_i^{norm} = {{{I_i} - \mu} \over \sigma} T=Rθ∘Fp∘Zs T = {R_\theta} \circ {F_p} \circ {Z_s} Ipre=TCLAHEIi−μ/σ,α|Ii∈I {I_{pre}} = \left\{{T\left({CLAHE\left({\left({{I_i} - \mu} \right)/\sigma,\alpha} \right)} \right)|{I_i} \in I} \right\} where I_i is the original image, μ and σ are the mean and standard deviation, R_θ applies rotation by θ, F_p performs flipping with probability p, Z_s applies zoom by scale s and α is the CLAHE clip limit. This preprocessing pipeline ensures uniform intensity, improved lesion visibility and robust feature learning for subsequent CNN-based DR classification. Table 1 shows the preprocessing steps in retinal disease.

Table 1:

Original image
Iinorm I_i^{norm}
Rθ
Fp
Zs

In this research, transfer learning utilizing pre-trained CNN models, such as ResNet50 and EfficientNet, extracts disease-specific characteristics from retinal fundus data. These models, trained on ImageNet, have extensive feature extraction capabilities that are fine-tuned on medical datasets. The models can detect retinal microaneurysms by freezing lower layers and adjusting upper layers to domain-specific properties. This method reduces training time, enhances accuracy and leverages learned visual hierarchies to diagnose diseases in multimodal medical imaging.

b.i

Transfer learning with pre-trained CNN backbones (ResNet-50, EfficientNet-B0)

The TL-Pretr-CNN system uses ImageNet-trained backbones (ResNet50, EfficientNet) to grade DR. High-resolution fundus images undergo preprocessing, followed by the application of frozen foundation layers to maintain universal edge and texture detectors. The final convolutional block and a bespoke softmax head are fine-tuned on retinal input to reduce training costs and prevent overfitting. The original feature hierarchy is maintained while a limited parameter subset is updated using a unified loss (cross-entropy plus L2 regularization). GAP and Grad-CAM improve clinician trust by compressing spatial maps into a compact vector for categorization and highlighting lesion locations. Let a color fundus image be I ∈ ℝ^224×224×3 after preprocessing. The components of a generic ImageNet network can be broken down into (4) fimageNet(I)=gbase(I;θbase)||ghead(.;θhead) {f_{imageNet(I)}} = {g_{base}}(I;{\theta _{base}})||{g_{head}}(.;{\theta _{head}})

In Equation (4), g_base denotes the stack of early and intermediate convolutional layers (e.g., ResBlocks in ResNet-50 or MBConv blocks in EfficientNet) that learn generic low-level features such as edges, textures, and shapes. g_head represents the high-level classification layers, including the global average pooling layer and the original 1,000-class softmax classifier.

a) Freezing the Generic Feature Extractor: To conserve universal visual filters, we freeze θ_base from from Eq. (5): (5) ∂L∂θbase=0⇒θbase*=θbaseImageNet {{\partial L} \over {\partial {\theta _{base}}}} = 0 \Rightarrow \theta _{base}^* = \theta _{base}^{ImageNet}

To freeze transfer learning, block modifications to the generic feature extractor's weights (θ_base). To preserve the generic visual features learned during pre-training, the parameters of the feature extraction backbone are frozen during training. This is achieved by setting the gradient of the training loss L with respect to the base network parameters θ_base to zero ∂L∂θbase=0. {{\partial L} \over {\partial {\theta _{base}}}} = 0.

After training, the optimal base parameters θbase* \theta _{base}^* , remain identical to the pre-trained ImageNet weights θbaseImageNet \theta _{base}^{ImageNet} . Here, θ_base denotes the parameters of the early and intermediate convolutional layers responsible for learning generic visual features such as edges, textures, and shapes. Freezing these layers allows the network to optimize only the task-specific classification layers, thereby reducing training time, minimizing the risk of overfitting on limited medical datasets, and leveraging robust feature representations learned from the large-scale ImageNet dataset.

b) DomainSpecific FineTuning: Utilize a custom DR classifier g_cls with parameters θ_cls and adjust the last convolutional block g_top (e.g., Conv5_x in ResNet-50) from Eq. (6): (6) F=gtopgbaseI;θbase*;θtop∈ℝ7×7×d F = {g_{top}}\left({{g_{base}}\left({I;\theta _{base}^*} \right);{\theta _{top}}} \right) \in {{\mathbb R}^{7 \times 7 \times d}}

A medical image I is processed through the transferlearning network to yield a high-level feature map. The inner function gbase represents the frozen, generic feature extractor whose parameters θbase* \theta _{base}^* are locked (equal to the ImageNet-trained values). It converts the input image into intermediate activations. These activations are then passed to gtop, the trainable adaptation block parameterized by θ_top is fine-tuned on domain-specific data to capture pathological cues. The final output F is a tensor of spatial dimension 7 × 7 with d channels, encapsulating rich, task-aware visual descriptors that feed subsequent pooling or classification layers. Applying GAP yields from Eq. (7): (7) z=GAP(F)=149∑h=17∑w=17Fh,w,∈ℝd z = GAP(F) = {1 \over {49}}\sum\limits_{h = 1}^7 {\sum\limits_{w = 1}^7 {{F_{h,w, \in {{\mathbb R}^d}}}}}

The d-dimensional activation vector at the spatial location (h,w) within the 7 × 7 feature map generated by the fine-tuned CNN is denoted by F_h,w ∈ R_d, where h and www represent the spatial coordinates and d is the number of feature channels. Summing all 49 location vectors and dividing by 49 yields the mean activation for each channel, keeping channel-wise semantics while reducing spatial dimensions. A compact d-length representation, z, contains the most important aggregated information from the full feature map. Reduced dimensionality prevents overfitting, integrates smoothly into fully connected or softmax layers and maintains the network's capacity to identify patterns across the image. The custom head maps z to the C DR grades (C = 3 in: noDR, mild, moderate) in Eq. (8): (8) y^=gcls(z;θcls)=Softmax(Wz+b),W∈ℝc×d {\hat y} = {g_{cls}}(z;{\theta _{cls}}) = Softmax(Wz + b),W \in {{\mathbb R}^{c \times d}} where W ∈ ℝ^c×d and b ∈ ℝ^c are learnable parameters. Using GAP, the channel-wise feature vector obtained is z ∈ ℝ, z=149∑h=17∑w=17Fh,w z = {1 \over {49}}\sum\limits_{h = 1}^7 {\sum\limits_{w = 1}^7 {{F_{h,w}}}} . The linear transformation Wz+b converts pooled characteristics into a C-dimensional logit vector, where c represents the number of disease classes (e.g., DR severity levels). The softmax function translates logits into a probability distribution y^ \hat y , where the ith component indicates the network's confidence in class i for the input image. The top layers (W and b) are trained end-to-end, while the GAP representation (z) ensures that spatially summarized, semantically rich information drives the classification choice.

The pipeline begins with preprocessed fundus photos fed into a frozen ImageNet backbone (ResNet-50 or EfficientNet-B0), as illustrated in Figure 2. The locked parameters θbaseImageNet \theta _{base}^{ImageNet} . As shown by the edge detection and feature enhancement responses, the frozen convolutional layers preserve edge and texture detectors learned during pre-training.

Figure 2:

TLPretrCNN workflow for multimodal retinal diagnosis. CNNs, convolutional neural networks; DR, diabetic retinopathy; GAP, global average pooling; TL-Pretr-CNN, transfer learning with pre-trained CNN models.

b.ii

Transfer learning optimization for medical images

The implied TL-Pretr-CNN, designed for combined MRI of the retina, is shown in the panel as it gradually improves visual evidence from generic edges to pathology-aware signs. It is illustrated in Figure 3. Four representations are sent to the network in the blue “Input Pattern” column: (a) the contrast-normalized image, (b) the edge map, (c) the fused pathological prior and (d) the Grad-CAM saliency seed. The most activated channels following each residual block (Conv2–Conv5) are shown in green columns, a–d. The ability of Conv2 (the frozen version of BaseNet) to respond to universal contours, such as optic disk rings and cranial outlines, shows that it has retained the primitives from ImageNet. As the partially trainable α head becomes more adaptive, Conv3 starts to specialize in tasks, retinal maps improve micro-aneurysm rims and MRI maps define ventricular borders. Fully trainable Conv4 ( θ top) enhances lesion-centric filters; activations with color coding identify tumor textures and hemorrhages, demonstrating domain transfer. The interpretability is validated when Conv5 feeds the custom softmax head; Grad-CAM overlays co-localize with clinically significant regions.

Figure 3:

Progressive feature evolution across TLPretrCNN residual stages. CNNs, convolutional neural networks; TLPretrCNN, transfer learning with pre-trained CNN models.

Hyperparameter selection and optimization

In the TL-Pretr-CNN framework, feature extraction and selection are implicitly learned by the network, rather than via explicit feature selection or weighting techniques. The key hyperparameters were carefully chosen to ensure stable training and optimal performance: a batch size of 32 balances GPU memory limits and gradient stability; a learning rate of 1 × 10⁻⁴ with cosine annealing ensures smooth convergence and improves final accuracy; the AdamW optimizer with weight decay 0.01 was selected for superior generalization; and 40 training epochs provided the best trade-off between performance and overfitting. The last three convolutional blocks of the backbone were fine-tuned to adapt pre-trained features to retinal pathology, while data augmentation—including CLAHE, ±20° rotations, random cropping and brightness/contrast adjustments—enhanced model generalization. Hyperparameters were empirically optimized using grid search and sensitivity analysis based on evaluation metrics such as clinical diagnostic accuracy (CDA), lesion localization performance (LLP), severity grading concordance (SGC) and Explainability Confidence Score (ECS).

This approach allows the CNN to automatically focus on the most informative retinal regions, contrasting with traditional feature selection methods that require manual scoring or statistical evaluation. Table 2 summarizes the selected hyperparameters and their corresponding values.

Table 2:

Hyperparameter	Selected value	Notes/optimization rationale
Batch size	32	Balances GPU memory and gradient stability
Learning rate	1 × 10−4 (cosine annealing)	Smooth convergence; avoids oscillation or plateauing
Optimizer	AdamW (weight decay = 0.01)	Strong generalization compared to Adam
Epochs	40	Trade-off between performance gain and overfitting
Fine-tuning depth	Last three convolutional blocks	Adapts pre-trained features to retinal pathology
Data augmentation	CLAHE, ±20° rotation, random crop, brightness/contrast	Improves model generalization and robustness by increasing the diversity of the training data and reducing overfitting.

CLAHE, contrast-limited adaptive histogram equalization.

Classification and prediction with custom head adds a dense (completely linked) layer to a CNN's feature extractor for medical classification. After extracting deep features with ResNet-50 or EfficientNet, a classification head comprising one or more dense layers and a softmax layer is employed. This setup allows the model to assign probabilities to predefined classes (e.g., DR severity or tumor presence). During network training, the cross-entropy loss optimizes the classification head weights to discriminate between disease states accurately.

In the TL-Pretr-CNN, a custom classification head predicts disease severity from retrieved features. After using deep CNNs such as ResNet50 or EfficientNet, the feature vector z_i ∈ ℝ^D. It is processed with a fully connected layer and softmax activation to generate class probabilities. Let W ∈ ℝ^D×C is weight matrix for the dense layer, b ∈ ℝ^c is bias vector and y^i∈ℝC {\hat y_i} \in {{\mathbb R}^C} is predicted probability for C classes. The prediction equation uses a softmax function to translate feature vectors z_i into class probabilities y^i {\hat y_i} . (9) y^i=SoftmaxWTzi+b=eWTzi+b∑j=1CeWTzi+bj {\hat y_i} = Softmax\left({{W^T}{z_i} + b} \right) = {{{e^{{W^T}{z_i} + b}}} \over {\sum\nolimits_{j = 1}^C {{e^{\left({{W^T}{z_i} + b} \right)j}}}}}

The weights W and bias b translate deep characteristics to C output classes. Softmax assures outputs add to 1, turning them into probabilities. The training objective employs cross-entropy loss L to penalize the difference between the predicted distributions y^i {\hat y_i} . (10) L=−∑i=1N∑C=1Cyi,clog(y^i,c) {\cal L} = - \sum\limits_{i = 1}^N {\sum\limits_{C = 1}^C {{y_{i,c}}\log ({{\hat y}_i},c)}} where y^ic {\hat y_i}c is ground truth (1 if image i belongs to class C, else 0) and N is the number of training samples. This classification mechanism enables multi-class prediction for DR (no DR, mild, moderate, severe) and binary classification. Evaluation metrics such as accuracy, AUC, sensitivity and specificity validate the model's clinical robustness and reliability. The classification head serves as the “decision cortex” of TL-Pretr-CNN. The dense layer projects general, pathology-aware features onto a task-specific axis, the C disease categories, after the backbone distills each fundus or MRI slice into a compact descriptor z_i (e.g., D = 2048 channels for ResNet-50).

Algorithm 1: TL-Pretr-CNN—Transfer Learning Framework for Multimodal Medical Image Analysis

Input: Dataset D = {If, IMRI}, Pre-trained CNN Model M (e.g., ResNet50, EfficientNet), Labels Y

Output: Predicted Labels y^ \hat y , Grad-CAM heatmaps, evaluation metrics

Step 1: Data Preparation

For each Img in D:

Preproc (resize, norm)

Augment (rot, flip)

Step 2: Load Pre-trained CNN Model

Load M (ImageNet): M = fInet

Freeze base layers: θbasefrozen {\boldsymbol {\theta}}_{{\boldsymbol {base}}}^{{\boldsymbol {frozen}}}

Add custom head: θtop

Step 3: Transfer Learning – Two-Phase Training

θ=θbasefrozen∪θtop {\boldsymbol {\theta}} = {\boldsymbol {\theta}}_{{\boldsymbol {base}}}^{{\boldsymbol {frozen}}} \cup {{\boldsymbol {\theta}}_{{\boldsymbol {top}}}}

For each epoch in training_epochs:

For each mini-batch (Ib, Yb) in the training set:

Forward Pass:

F = FInet (Ib) # Feature extraction

y^=Ffine tune(F)# Classification head \matrix{{\hat y = {F_{fine\,tune(F)}}} \hfill & {\# \,{\rm{Classification}}\,{\rm{head}}} \hfill \cr}

Compute Loss: L(y^,yb) {\cal L}(\hat y,{y_b})

Backward Pass:

update θtop (fine-tune top layers)

Step 4: Deep Feature Extraction with GAP

For each input I_pre in the test set:

Fi = CNN_intermediate_layer(Ipre)

zi=GAP(Fi)=1H′W′∑h=1H′∑w=1W′Fih,w,: {z_i} = GAP({F_i}) = {1 \over {H{'}W{'}}}\sum\limits_{h = 1}^{H{'}} {\sum\limits_{w = 1}^{W{'}} {{F_i}\left[ {h,w,:} \right]}}

#Apply Global Average Pooling:

Step 5: Model Evaluation

Predict on test set: y^test=model(Itest) {\hat y_{test}} = {\rm{model}}({I_{test}})

Compute metrics:

Accuracy = correct/total

AUC = area under ROC curve

Sensitivity = TP / (TP + FN)

Step 6: Explainability via Grad-CAM

For y^ \hat y in Y^_test \hat Y\_{\rm{test}} :

Gen GCAM from the final conv

Overlay GCAM on the original image

Step 7: Output Integration

If eval_m ≥ th:

Intg mdl → TH_DSS

Else:

Retrn y^ \hat y , Ptune

Figure 4 presents the multi-class Grad-CAM visualization for three diagnostic categories: (A) Healthy, (B) Mild Pathology, and (C) Moderate/Severe Pathology. Deep neural networks are sometimes referred to as “black boxes,” making it difficult to trust their fore-casts without a thorough understanding. Grad-CAM generates visual heatmaps to highlight which medical picture regions contributed more to the prediction. This enhances model interpretability and facilitates doctors' validation of AI-assisted diagnoses.

Figure 4:

Multi-class GradCAM visualization for three diagnostic categories (A) Healthy, (B) Mild pathology, (C) Moderate/severe pathology. GradCAM, gradient-weighted class activation mapping.

The grid compares the input fundus images (column i) with four visual explanation stages (columns ii–v). Rows A–C represent the three diagnostic categories: (A) Healthy, (B) Mild pathology, and (C) Moderate/Severe pathology, allowing visualization of the model's attention across different disease severities. Column ii presents low-level saliency heat maps that capture broad anatomical contours learned by the early CNN layers. Column iii shows pixel-importance masks, where mid-layer activations highlight coarse lesion regions such as hazy edema patches or hypo-intense areas. Column iv displays class-specific Grad-CAM visualizations, illustrating how the ResNet-50/EfficientNet-based model focuses on clinically relevant retinal abnormalities, including vascular structures and lesion regions. Column v presents binary relevance overlays that identify the top 20% of the most influential pixels for clinical inspection using Grad-CAM.

This research utilized an Ubuntu 22.04 machine equipped with dual NVIDIA RTX A6000 GPUs (each with 48 GB of memory), an AMD Threadripper 3990X CPU (64 cores) and 256 GB of RAM for experimentation. Python 3.10, PyTorch 2.1, TorchVision 0.16 and MONAI 1.3 were utilized for medical imaging. Preprocessing and visualization were done with OpenCV 4.8, Albumentations 1.3 and Matplotlib 3.8. Performance metrics were calculated with scikit-learn 1.4. ImageNet-pre-trained ResNet-50 and EfficientNet-B0 backbones were used for transfer learning with fine-tuning on deeper layers. The models were trained with the AdamW optimizer, cosine annealing, 32 batches and mixed precision across 40 epochs. Setting fixed random seeds and tracking training logs and Grad-CAM visuals with Weights & Biases provided reproducibility.

CDA compares model predictions to medical diagnoses. By identifying illness prevalence, severity and type, CDA prioritizes clinical relevance over generic correctness. It utilizes expert-verified labels and the seriousness of misclassification to determine diagnostic value. A high CDA indicates that the AI system matches clinician performance and is more likely to be trusted for real-world healthcare deployment, notably for DR disease detection.

The CDA of four models—IDRiD-Net, DeepMedic, EfficientNet + Grad-CAM Hybrid and the proposed TL-Pretr-CNN—for DR across several disease severity levels is shown on the x-axis, which shows categorical disease severity stages (e.g., no DR, mild, moderate, severe), and the y-axis provides CDA values from 0.0 to 1.0. TL-Pretr-CNN CDA is greater across all categories (Figure 5), with better alignment with expert-verified clinical diagnoses. It has near-perfect accuracy for “no DR” cases and good accuracy in complex categories, such as “severe DR.” Older models, such as IDRiD-Net and DeepMedic, had more variability and lower accuracy at higher severity levels. TL-Pretr-CNN's robustness, clinical value and interpretability in real-world medical applications are confirmed.

Figure 5:

CDA comparison across disease severity levels of DR. CDA, clinical diagnostic accuracy; DR, diabetic retinopathy; Grad-CAM, gradient-weighted class activation mapping; IDRiD-Net, improved deep retinal intelligence detector network; TL-Pretr-CNN, transfer learning with pre-trained CNN models.

Using visualization tools such as Grad-CAM, LLP evaluates the degree to which a model accurately highlights problematic areas. A comparison is made between the heatmaps generated by the model and the ground-truth lesion masks or annotations, and the overlap is computed using metrics such as inter-section over union (IoU). The presence of a high LLP indicates that the model correctly classifies and concentrates on medically significant areas. It is a crucial step in providing clinical transparency and determining whether or not the model's attention is aligned with disease-affected tissues. LLP evaluates how accurately the Grad-CAM or similar visualizations highlight disease-relevant regions.

Within the context of performance evaluation, LLP introduces an anatomical fidelity dimension, which helps ensure that the model is both interpretable and clinically applicable.

IoU measures are used to evaluate the lesion localization performance (LLPr) scores of DL models, utilizing Grad-CAM heatmaps for disease localization. Highlighting clinically significant regions of interest (ROIs) in input images ensures that models are accurate in diagnosis and physically interpretable. Figure 6 shows LLP scores for IDRiD-Net, DeepMedic, EfficientNet + Grad-CAM and TL-Pretr-CNN.

Figure 6:

LLP evaluation plot. Grad-CAM, gradient-weighted class activation mapping; IDRiD-Net, improved deep retinal intelligence detector network; TL-Pretr-CNN, transfer learning with pre-trained CNN models.

In SGC, the degree of agreement between the projected severity grades of the disease and the annotations made by experts is measured. Weighted kappa or ordinal classification accuracy is utilized to consider the degree to which an incorrect prediction is near the correct class (e.g., moderate vs severe DR). In conditions such as DR, treatment options are based on the severity, which is essential. When the SGC is high, it ensures that the model can accurately stratify patients according to their risk level, which aids in triage and resource prioritization. It supports applying the concept in healthcare systems where severity-driven decision-making is crucial. ECS demonstrates the accuracy of the model's highlighted regions in comparison to clinician-verified ROIs.

The SGC for different DL models across DR severity levels is shown in Figure 7. SGC measures model-predicted severity grades' reliability against expert ophthalmologist annotations using Quadratic Weighted Kappa. This metric measures grade agreement and disagreement, making it suited for ordinal, risk-stratified medical classification tasks. In mild to severe DR, TL-Pretr-CNN outperforms IDRiD-Net, DeepMedic and EfficientNet + Grad-CAM suggests a better assessment of sickness severity, which could improve clinical patient triage and therapy prioritization.

Figure 7:

SGC line chart across disease grades. DR, diabetic retinopathy; SGC, severity grading concordance.

The anatomical agreement between model-generated attention maps and clinician-annotated regions was assessed using the ECS. Equation (16) provides the mathematical definition of the ECS, where H_model denotes the Grad-CAM heatmap generated by the proposed model, and H_expert represents the expert-annotated lesion localization map provided by clinical specialists. The ECS quantifies the pixel-wise correspondence between these two heatmaps, with higher values indicating stronger agreement between the model's visual explanations and expert annotations.

In Equation (16), H_model ⊙ H_expert denotes the element-wise multiplication (pixel-wise dot product) between the model-generated Grad-CAM heatmap and the expert-annotated lesion localization map. The terms || H_model ||₂ and || H_expert || represent the L2 norms of the corresponding heatmaps, which are used to normalize the similarity score.

ECS measures a model's ability to provide an explanation for its decisions that is consistent with clinical reasoning. The accuracy (probability) of the forecasts is combined with the clarity (grad-CAM heatmaps) and consistency. For ECS to work, models must generate consistent activation maps that pinpoint areas where disease is most prevalent. This score increases confidence in AI conclusions by confirming that the medical expert can understand both the prediction and the explanation. An accurate and precise model, along with a high ECS, is essential for clinical acceptance and regulatory approval. SGC measures the agreement between predicted and actual disease severity levels (e.g., no DR, mild, moderate).

Figure 8 shows the ECS plots of IDRiD-Net, DeepMedic, EfficientNet + Grad-CAM and TL-Pretr-CNN DL models for retinal sickness diagnosis. The x-axis shows models, and the y-axis shows their ECS ratings ranging from 0.0 to 1.0. ECS utilizes the visual alignment of model-generated Grad-CAM heatmaps and expert-annotated lesion locations to evaluate how effectively a model approximates clinical reasoning by focusing on diagnostically relevant areas. Equation (16) defines the ECS, which quantifies the agreement between the Grad-CAM heatmap intensity distribution and the expert-annotated lesion localization map. The score is computed using the pixel-wise dot product between the Grad-CAM saliency map and the corresponding expert annotation, while the L2 norm is used to normalize the magnitude of each heatmap distribution. Higher ECS values indicate better alignment between the model's visual explanations and clinically annotated lesion regions, thereby improving the interpretability of the predictions. The proposed TL-Pretr-CNN achieved the highest ECS score (0.89), demonstrating strong correspondence between the generated visual explanations and expert annotations. In contrast, IDRiD-Net obtained the lowest ECS score (0.45), indicating poor anatomical agreement and reduced interpretability. Additionally, Cohen's Kappa (κ) coefficient was employed to evaluate the agreement between the predicted and reference severity grades, thereby confirming the reliability of the severity classification results.

Figure 8:

ECS evaluation plot. ECS, explainability confidence score; Grad-CAM, gradient-weighted class activation mapping; IDRiD-Net, improved deep retinal intelligence detector network; TL-Pretr-CNN, transfer learning with pre-trained CNN models.

The TL-Pretr-CNN framework was evaluated against baseline models—IDRiD-Net, DeepMedic and EfficientNet + Grad-CAM—on the test set using multiple metrics: Accuracy, LLP, SGC, ECS, sensitivity, specificity and AUC.

To assess whether the observed improvements were statistically meaningful, paired t-tests were conducted across five independent runs with different random seeds. For each metric, the mean and standard deviation were computed and p-values were calculated. The TL-Pretr-CNN showed statistically significant improvements (p < 0.05) in accuracy, LLP and SGC compared to all baseline models across all DR severity levels, indicating that the performance gains are robust and not due to random variation.

Inference time was measured per image on a system with an NVIDIA A100 GPU (48 GB) and an Intel Xeon 32-core CPU, using a batch size of 32. TL-Pretr-CNN achieved an average inference time of ~12 ms per image, demonstrating suitability for real-time teleophthalmology applications. For comparison, IDRiD-Net and DeepMedic were slightly faster (~8–10 ms/image) due to their smaller architectures, whereas EfficientNet + Grad-CAM required ~14 ms/image because of Grad-CAM computations. Table 3 shows the comparative analysis.

All reported metrics for TL-Pretr-CNN and baseline models were averaged over five independent training runs, each initialized with a different random seed. This approach captures the natural variability arising from stochastic gradient descent, weight initialization and mini-batch sampling, ensuring that performance estimates are robust. The mean and standard deviation are reported for all key metrics (accuracy, LLP, SGC, ECS, sensitivity, specificity, AUC), providing a measure of variability and reliability.

Evaluation was conducted on a held-out test set (15% of the dataset), which was not used during training or validation. Stratified sampling ensured proportional representation of all DR severity classes. While explicit k-fold cross-validation was not performed, the validation set (15%) was used to tune hyperparameters and monitor overfitting. The held-out test set, therefore, provides an unbiased assessment of generalization performance.

While the TL-Pretr-CNN framework demonstrates high accuracy, LLP and SGC, several limitations remain. First, generalizability may be affected by the fact that the model was trained and evaluated on a single, publicly available retinal fundus dataset; performance on images from different populations, imaging devices or clinical settings may vary. Second, although inference time (~12 ms per image with batch size 32 on an A100 GPU) is suitable for teleophthalmology, the computational cost and GPU requirements could limit deployment in low-resource or mobile settings. Third, extreme DR cases with rare pathological patterns may still be misclassified due to class imbalance in the dataset.

To address these limitations, future work includes prospective clinical validation in hospital or teleophthalmology environments, evaluating the system on diverse patient cohorts and optimizing the model for edge deployment. This will assess real-world effectiveness, reliability and integration with decision support workflows for ophthalmologists.