Retinal Vessel Segmentation and Lesion Detection for Diabetic Retinopathy Diagnosis Using Context-Aware Deep Neural Networks on SPECTRAL OCT Data

Diabetic Retinopathy (DR) has been examined through a variety of computational methods, each with unique benefits and drawbacks. By using features extracted from retinal images, traditional machine learning techniques like Random Forests (RF) [7] and Support Vector Machines (SVM) [8] have been used for DR classification. These methods work well for conditions that are clearly defined and have relatively small datasets.

Convolutional Neural Networks (CNNs), such as VGG16 and ResNet, have been extensively utilized for DR diagnosis and segmentation tasks since the development of deep learning [9, 10]. By autonomously learning hierarchical features from data, these models increase the precision and resilience of lesion detection and retinal vascular segmentation. Despite being successful, the models' high computational demand prevents real-time clinical implementation, and they frequently require sizable annotated datasets for efficient training. Hybrid methods combining CNNs with other techniques, such as U-Net for segmentation and Fully Convolutional Networks (FCNs) for pixel-wise classification, have further improved segmentation performance. These architectures excel at capturing spatial and contextual information, enabling precise vessel segmentation and lesion detection [11]. However, its performance degrades in the presence of artifacts in retinal images.

Advanced techniques have been developed to enhance DR detection and segmentation in addition to conventional Machine Learning (ML) and early deep learning techniques. To overcome the difficulties presented by the scarcity of annotated datasets, Generative Adversarial Networks (GANs) have been utilized to improve image quality and carry out data augmentation [12]. GANs also enable domain adaptation, improving generalization across diverse imaging modalities. However, they are computationally intensive and challenging to train, often requiring careful hyperparameter tuning. Graph Neural Networks (GNNs) have been explored for capturing relational and structural information in retinal images, particularly for vessel segmentation tasks [13]. These models use graph representations to model the connectivity of retinal vasculature, resulting in improved segmentation accuracy. However, GNNs are limited by their complexity and dependency on accurate graph construction.

Transformer-based architectures, such as Vision Transformers (ViTs), have recently been adapted for DR diagnosis. These models use self-attention mechanisms to focus on important regions in retinal images, offering improved interpretability and accuracy [14]. However, their performance often relies on pretraining with large-scale datasets, which may not always be available. Ensemble learning approaches, combining multiple models like CNNs, U-Nets, and decision trees, have also shown the potential to improve robustness and accuracy for DR detection. Ensemble techniques reduce the drawbacks of single-model approaches by utilizing the advantages of individual models [15]. However, these techniques are computationally intensive and challenging to use in clinical settings in real time.

Transfer learning speeds up model construction and eliminates the requirement for large amounts of labeled data. Still, these models frequently need to be adjusted to fit DR datasets and might not adequately capture domain-specific properties. Because CapsNets can record spatial hierarchies and interactions between visual features, they have been explored for DR detection [16]. Unlike CNNs, which may lose spatial information during pooling, CapsNets preserve positional data, improving accuracy in lesion detection. However, the computational complexity and extended training time limit their practical application in real-time settings.

Temporal variations in sequential imaging data have been analyzed for DR progression analysis using models based on Recurrent Neural Networks (RNNs), such as Long Short-Term Memory (LSTM) networks. These methods are suitable for longitudinal studies but are less suited for single-image analyses and require consistent follow-up data [17]. Multitask Learning (MTL) frameworks simultaneously perform DR diagnosis and segmentation tasks by sharing a common feature extractor across tasks. This approach improves computational efficiency and reduces overfitting [18]. However, balancing multiple loss functions can be challenging, and task interference may degrade overall performance. DR datasets with few annotations can benefit from Few-Shot Learning (FSL) algorithms, which overcome data scarcity by learning to generalize from a small number of labeled instances [19]. While promising, FSL models are sensitive to variations in data quality and often require carefully curated support sets for training. Active Learning (AL) strategies involve iterative model training with user feedback to label the most informative samples. This reduces the annotation while maintaining model performance [20, 21]. However, the approach relies heavily on the availability of expert annotators and can be time-intensive for large datasets.

To detect diabetic retinopathy, many methods for segmenting retinal blood vessels and lesions have been reported. These methods are differentiated using the proposed technique to provide a comprehensive review of this wide range of approaches.

The proposed approach for diagnosing diabetic retinopathy utilizing 3D-Spectral Optical Coherence Tomography (OCT) images is shown in Figure 2. First, 3D-Spectral OCT images are given. Next, a preprocessing phase employing CLAHE is performed to improve image contrast and provide focus on subtle retinal features. Following preprocessing, the images are sent into a CEDNN for feature extraction, producing two essential outputs: lesion identification and retinal vascular segmentation.

image004.png

Figure 2. The proposed framework for DR diagnosis

Lesion detection looks for anomalies such as microaneurysms, hemorrhages, and exudates, which are significant markers of DR, whereas retinal vascular segmentation concentrates on recognizing blood vessel architecture. The final DR diagnosis is the result of an analysis based on the combination of these results. This pipeline ensures precise localization and classification of DR-related features, facilitating accurate and efficient diagnosis.

4.1 Preprocessing

The image has been divided into tiles, which are little contextual areas that do not overlap. Let $T\times T$ be the size of each contextual region and $I\left( {{x}_{1}},{{y}_{1}} \right)$ be the intensity of the pixel of the original image at position $\left( {{x}_{1}},{{y}_{1}} \right)$. Determine the histogram H(i) for every tile using Eq. (1), where i$\in${0,1,…,L−1} represents the intensity levels.

$H\left( i \right)=\underset{x,y\in title}{\mathop \sum }\,\delta \left( I\left( {{x}_{1}},{{y}_{1}} \right)-i \right)$                (1)

Here $\delta \left( k \right)$ is the Kronecker delta function.

If $I\left( {{x}_{1}},{{y}_{1}} \right)=i~then~\delta \left( I\left( {{x}_{1}},{{y}_{1}} \right))-i=1 \right)~$contributing to the histogram bin $H\left( i \right)$. Otherwise, the contribution is 0. This ensures that the histogram is built by counting the exact matches of pixel intensities within each contextual region (tile). To prevent over-amplification of noise, clip the histogram values at a predefined threshold $C$ as shown in Eq. (2).

${{H}_{clip}}\left( i \right)=\text{min}\left( H\left( i \right),C \right)$                 (2)

The excess pixels above the threshold are redistributed across the histogram bins. The clipped histogram is normalized to compute and display the local Cumulative Distribution Function (or CDF) for every tile, as shown in Eq. (3).

$CDF\left( i \right)=\frac{\mathop{\sum }_{j=0}^{i}{{H}_{clip}}\left( j \right)}{\mathop{\sum }_{j=0}^{L-1}{{H}_{clip}}\left( j \right)}$                 (3)

The CDF maps the intensity values to its adjusted levels. The pixel intensity is transformed using the local CDF as shown in Eq. (4):

${I}'\left( {{x}_{1}},{{y}_{1}} \right)=\text{min}+CDF(I\left( {{x}_{1}},{{y}_{1}} \right)\times \left( max-\text{min} \right)$                (4)

where, min and max represent the image's lowest and highest intensity values, respectively. To avoid artificial boundaries between tiles, bilinear interpolation is applied to smooth the transition of pixel intensities.

${I}''\left( {{x}_{1}},{{y}_{1}} \right)=interpolate\left( {I}'\left( {{x}_{1}},{{y}_{1}} \right),~neighboring~tiles \right)$                  (5)

As shown in Figure 3, for DR 3D-OCT images, CLAHE enhances contrast in areas with subtle variations caused by diabetic pathologies, such as microaneurysms and exudates, by emphasizing local differences without over-enhancing noise. While CLAHE is more commonly applied in fundus photography, its use in OCT preprocessing was quantitatively evaluated through an ablation study.

tu_pian_1.jpg

(a)

tu_pian_2.png

(b)

Figure 3. Preprocessing: (a) Input OCT image [28], (b) Preprocessed output image

Applying CLAHE improved the Dice score by +2.4% for vessel segmentation compared to simple intensity normalization, confirming its benefit for enhancing fine microvascular structures. At the same time, acknowledge that CLAHE may risk altering OCT reflectivity patterns, which could distort anatomical interpretation. To mitigate this, CLAHE parameters of tile size and clip limit were optimized to balance enhancement with noise suppression.

4.2 CEDNN for segmentation

A CEDNN is a deep learning architecture intended for DR diagnostic image segmentation and feature extraction. Although the dataset originates from 3D SPECTRAL OCT volumes, the proposed model processes data at the level of 2D B-scan slices extracted from these volumes. In this context, 3D refers to the acquisition modality, while segmentation is performed slice-by-slice using CEDNN. Each B-scan is treated as a grayscale image with input dimensions of [Height×Width×Channel], where Channel=1. This design choice avoids the computational load of 3D convolutions while capturing volumetric information across multiple slices during evaluation.

The encoder captures high-level characteristics and compresses input data, while the decoder reconstructs the input's spatial structure for segmentation. These two components define the network. The CEDNN stands out compared to other models for DR diagnosis due to its higher balance between feature extraction and spatial detail preservation.

While traditional CNNs like VGG or ResNet are excellent for global feature extraction, they often lose spatial resolution, which is vital for localizing small lesions such as microaneurysms and hemorrhages.

Figure 4 illustrates the architecture of a CEDNN tailored for 3D-OCT image segmentation. The process begins with the feature encoder, which extracts spatial features using a series of convolutional layers with 3×3 convolution to capture local spatial information. The encoded features are then passed to the context extractor, which applies further convolutions and operations like skip connections and feature aggregation to enhance contextual understanding across scales. Next, the processed features are passed to the feature decoder, which reconstructs the spatial details of the segmented regions. The decoder includes components like a 1×1 convolution for dimensionality reduction, 3×3 deconvolutions with upsampling layers, and additional 3×3 convolutions for refining the segmentation output. The segmentation accuracy of complicated structures in 3D-OCT images is improved by skip connections between the encoder and decoder, which guarantee that fine-grained information from the encoder is preserved throughout reconstruction. This architecture is designed to achieve high accuracy in identifying regions of interest within 3D-OCT images by utilizing hierarchical feature extraction and reconstruction.

image007.png

Figure 4. The proposed architecture of CEDNN for 3D-OCT image segmentation

CEDNN's encoder-decoder architecture mitigates this limitation by combining downsampling in the encoder for robust feature learning with upsampling in the decoder for precise spatial reconstruction. Compared to Fully Convolutional Networks (FCNs) or U-Net, CEDNN offers enhanced scalability and flexibility, allowing it to process high-dimensional retinal images with greater efficiency. CEDNN is computationally efficient and well-suited for DR classification and segmentation tasks with limited data. The encoder component efficiently detects patterns linked to DR, including microaneurysms, hemorrhages, and exudates, by capturing hierarchical feature representations using convolutional layers. To precisely localize problematic areas, the decoder simultaneously reconstructs fine-grained spatial features using upsampling and deconvolution layers. This architectural design ensures the model retains spatial resolution for DR image analysis while capturing global contextual information. Furthermore, CEDNN is very effective for lesion segmentation and DR severity grading due to its multi-scale feature extraction and effective pixel-wise classification, which addresses the difficulties of lesion size and appearance variability in retinal fundus images.

Unlike standard encoder–decoder networks such as U-Net, the proposed CEDNN introduces a dual-context encoding module. This module integrates:

• Local context encoding through residual 3×3 convolutions, which preserve fine lesion details such as microaneurysms and vessel junctions.

• Global context encoding through dilated convolutions and multi-scale aggregation, which capture long-range vascular topology across OCT volumes.

The outputs of these two streams are mathematically aggregated to encode both fine-grained and large-scale retinal features simultaneously. By explicitly modeling global retinal layer continuity in addition to localized lesion detail, CEDNN provides a context-aware representation that standard U-Net and SegNet architectures do not achieve.

4.2.1 Encoder feature extraction

The encoder uses pooling and convolutional processes to compress the input image I into a lower-dimensional latent representation. The convolution operation extracts feature maps as shown in Eq. (6).

${{F}_{l}}=\sigma \left( {{W}_{l}}*{{F}_{l-1}}+{{b}_{l}} \right)$                     (6)

where, the input feature map to layer l is ${{F}_{l-1}},~$indicates the convolution process, while ${{W}_{l}}~$and ${{b}_{l}}$ represent the learnable weights and biases of layer l. The activation function is represented by σ, and ReLU is $ReLU\left( \sigma \left( x \right)=\max \left( 0,x \right) \right)$.

Eq. (7) illustrates how pooling layers preserve significant properties while reducing the spatial dimensions.

$F_{l}^{pooled}=Pool\left( {{F}_{l}} \right)$                (7)

Common pooling methods include max pooling as shown in Eq. (8).

$F_{l}^{pooled}\left( i,j \right)=\underset{p,q\in R}{\mathop{\max }}\,{{F}_{l}}\left( p,q \right)$                    (8)

where, R is a small window of 2×2 over which the maximum value is taken. This module extracts low-level features from the input 3D-OCT images using multiple 3×3 convolutional layers. These layers capture spatial details, such as the boundaries of retinal layers and abnormal structures like microaneurysms or drusen. Skip connections from the encoder are established to pass significant spatial information to the decoder, preserving fine details for accurate segmentation.

4.2.2 Context extractor

Accurate segmentation in 3D-OCT image processing depends on the model's capacity to collect local and global context, which is improved by the context extractor. The extracted features are processed to enhance the contextual understanding across different regions of the 3D-OCT image. Operations like residual connections improve learning by enabling the reuse of features and stabilizing the gradient flow. The context extractor uses convolutional layers to capture spatial hierarchies. The convolution operation is defined as shown in Eq. (9).

$y_{i, j, k}=\sum_{m=1}^M \sum_{n=1}^N x_{i+m-1, j+n-1, k} \cdot w_{m, n, k}+b_k$                    (9)

where,

• ${{x}_{i,j,k}}$: Input feature map at position (i,j) and channel k.
• ${{w}_{m,n,k}}$: Convolutional kernel of size M×N (3×3).
• ${{b}_{k}}$: Bias term for the k^th channel.
• ${{y}_{i,j,k}}$: Output feature map after applying the convolution.

These layers capture local spatial details in the input feature map. By stacking multiple convolutions, the network aggregates multi-scale context, which is particularly useful in detecting subtle patterns in OCT images. Deeper networks are made possible by residual connections, which also enhance gradient flow. Eq. (10), which provides the residual operation.

$Y=F\left( X,\left\{ W \right\} \right)+X$                    (10)

where,

• X: Input feature map.
• F (X, {W}): Non-linear transformation of the input, such as convolutions.
• {W}: Learnable weights of the transformation.
• Y: Output feature map after residual addition.

Training is stabilized, and vanishing gradients are avoided because of residual connections, which enable the model to learn the variation between the input and output. Deep contextual feature extraction from 3D-OCT images requires it. Without adding more parameters, dilated atrous convolutions improve the receptive field. Eq. (11) provides a definition for the operation.

$y\left[ i \right]=\underset{k=1}{\overset{K}{\mathop \sum }}\,x[i+r·k]·w\left[ k \right]$                       (11)

where,

• r: Dilation rate of r=2 doubles the receptive field.
• K: Kernel size.
• $x\left[ i \right]$: Input feature map.
• $w\left[ k \right]$: Kernel weights.

Dilated convolutions capture global context, which is essential for identifying large structures or lesions in 3D-OCT images, such as retinal detachment. The extracted features are aggregated using either concatenation or element-wise addition to combine information across multiple scales, as shown in Eq. (12):

${{Y}_{agg}}={{Y}_{1}}\oplus{{Y}_{2}}\oplus\ldots \oplus{{Y}_{n}}$                  (12)

where,

• $\oplus$: Aggregation operation (concatenation or addition).
• ${{Y}_{1}},$ ${{Y}_{2}},\ldots ,~{{Y}_{n}}$: Feature maps from different layers.

Aggregation combines fine-grained details with broader contextual information, crucial for distinguishing overlapping structures in OCT images.

4.2.3 Decoder spatial reconstruction

The decoder reconstructs the high-resolution output of the segmentation mask from the compressed feature map using upsampling and convolution. The decoder uses upsampling operations to increase spatial dimensions, often using transposed convolutions, as shown in Eq. (13).

$F_{l}^{up}=\sigma \left( W_{l}^{T}*{{F}_{l+1}}+{{b}_{l}} \right)$                    (13)

Here, $W_{l}^{T}~$represents the weights for transposed convolution, also called deconvolution. Skip connections connect the encoder and decoder layers to recover spatial features lost during downsampling, as indicated in Eq. (14).

$F_{l}^{skip}=F_{l}^{encoder}+F_{l}^{decoder}$                       (14)

This ensures the final output combines high-level and low-level features. This module reconstructs the spatial details for segmentation. It uses 1×1 convolutions to reduce feature dimensions, 3×3 deconvolutions for upsampling, and additional 3×3 convolutions to refine the segmentation output. For the purpose of defining fine structures in 3D OCT images, the skip connections between the encoder and decoder guarantee that the high-resolution information from the encoder is included throughout reconstruction.

4.2.4 Output layer

The segmentation mask or classification map is generated by the decoder's last layer. As shown in Eq. (15), segmentation is accomplished by using a softmax activation function pixel-by-pixel to assign each pixel to one of the C classes.

${{P}_{c}}\left( x,y \right)=\frac{\text{exp}\left( {{F}_{L}}\left( x,y,c \right) \right)}{\mathop{\sum }_{k=1}^{C}\text{exp}({{F}_{L}}\left( x,y,k \right)}$                    (15)

where, the probability of class c at pixel $\left( x,y \right)$ is denoted by ${{P}_{c}}\left( x,y \right)$.The output of the last layer for class c is ${{F}_{L}}\left( x,y,c \right)$.

In DR diagnosis, CEDNN enables precise retinal vessel segmentation and lesion detection, vital for identifying pathologies and classifying disease severity. The lesion detection using CEDNN is designed as a pixel-wise segmentation module that produces masks for clinically relevant lesion categories, including microaneurysms, hemorrhages, and exudates. This differs from bounding-box detection or image-level classification, as each lesion pixel is explicitly labeled. The lesion segmentation maps are then aggregated with vessel segmentation maps to provide complementary biomarkers, which are subsequently used in the diagnostic stage. Figure 2 has been updated to illustrate this dual-output pipeline more clearly.

4.2.5 Diagnostic classification

The outputs of the vessel segmentation and lesion detection branches are not used directly for grading. Instead, quantitative biomarkers are extracted from these segmentation maps, including vessel density, branching patterns, lesion count, and lesion area. These biomarkers serve as structured input features for a random forest classifier, which assigns a diabetic retinopathy severity grade according to standard clinical categories. This additional classification stage ensures that the model provides a clinically interpretable output rather than raw segmentation maps, linking image analysis to diagnostic decision-making.

The proposed DR 3D-OCT image segmentation method is assessed based on its ability to properly segment retinal features such as blood vessels, lesions, and other DR-related anomalies. A thorough evaluation of the algorithm's segmentation accuracy is provided by the evaluation metrics utilized for performance analysis, which comprise True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) values. The True Positive (TP) value shows how many pixels in the segmented and ground truth images were accurately classified as lesions or vasculature. The algorithm's capacity to identify the DR-affected regions is shown in higher TP values. The number of pixels in the segmented picture and the ground truth that are accurately identified as non-lesions or non-vessels is known as the True Negative (TN). This indicates the model's capacity to properly discriminate between normal and damaged retinal areas. Pixels in the segmented image that are mistakenly identified as lesions or vessels but do not match any lesion or vessel in the ground truth are known as false positives (FP).

Better segmentation accuracy, which reduces false detections, is indicated by a lower FP value. The number of pixels in the ground truth that represent lesions or vasculature but are not picked up in the segmented output is known as the False Negative (FN). To make sure the algorithm doesn't overlook essential DR-related characteristics like microaneurysms, exudates, or hemorrhages, FN must be kept to a minimum. The performance metrics for DR segmentation in 3D-OCT images using the proposed CEDNN model are shown in Table 2. With a 99.3% accuracy, 99.4% sensitivity, and 99% specificity, the approach performs well. The F1-score, which provides an acceptable balance between precision and recall, is 99.3%, while the algorithm's precision is 99.2%. Additionally, the proposed technique's resilience and dependability in precisely segmenting DR lesions and blood vessels are demonstrated by the estimated 95% Confidence Interval (CI) for accuracy, which ranges from 98.8% to 99.8%.

Figure 8 illustrates the model accuracy of 99.3% indicating correct predictions for positive and negative cases. The model also has a low error rate, with only 350 FP and 350 FN. With precision, recall, and accuracy closely matched at 99.3%, these findings demonstrate the model's high precision in differentiating between positive and negative instances, showing efficient detection and minimal misclassification.

image022.png

Figure 8. Confusion matrix illustrating the classification performance of the proposed model

As shown in Figure 9, the ROC curve and AUC serve as essential measures for assessing the way DR diagnostic classification algorithms work. The trade-off between the True Positive Rate (TPR) and the False Positive Rate (FPR) across various categorization thresholds can be illustrated by the ROC curve. AUC measures the model's overall performance; higher classification accuracy is indicated by numbers nearer 1. In DR images, the CEDNN model successfully identified lesions, exudates, and microaneurysms while separating healthy from afflicted regions with a high AUC of 99.6%. This result demonstrates the extent to which the CEDNN model can recognize and segment DR, which is essential for precise diagnosis and treatment planning.

image023.png

Figure 9. The AU-ROC of the proposed approach

High performance was attained by the proposed CEDNN, which included multiple measures to protect against data leaks. At the patient level, training, validation, and test sets have been split, and Table 2 presents the findings along with the 95% Confidence Intervals (CIs). The high performance may in part reflect dataset homogeneity and the influence of CLAHE preprocessing. However, risks of overfitting remain, particularly in cross-dataset generalization. To ensure fair evaluation, baseline models including DR-VNet, U-Net, ResNet+, and VGG16 CNN were re-implemented and trained on the same dataset with an identical train/test split as CEDNN. This design isolates the contribution of architectural differences rather than dataset variations. For models that could not be fully reproduced due to unavailable implementation details, results from the original publications were used and are marked in Table 3 to avoid misinterpretation.

Table 2. Evaluation metrics [15] and performance results of the proposed CEDNN for DR 3D-OCT ımage segmentation

Table 3. Comparative analysis of segmentation results obtained using CEDNN and state-of-the-art approaches [27-30] for DR detection

The performance of various cutting-edge DR segmentation methods, such as DR-VNet [27], Multiple U-Nets and SegNets [28], DR-ResNet+ [29], and VGG16-based CNN [30], is compared with that of the CEDNN approach in Table 3. With the highest levels of accuracy, sensitivity, and specificity, the CEDNN approach continuously outperforms the other approaches on all parameters. CEDNN's AUC results demonstrate its resilience and capacity to capture the intricate architecture of retinal blood vessels and diseases. DR-VNet [27] works well and produces competitive results, but it still falls short of CEDNN in terms of accuracy, precision, and F1-score. This indicates that CEDNN's 3D segmentation and context encoding methods yield more accurate results, especially for minor lesions and blood vessels. The multiple U-Nets and SegNets method exhibits lower performance, in precision and sensitivity, suggesting that it struggles with accurately segmenting fine details. Similarly, DR-ResNet+ [29] and VGG16-based CNN [30] achieve good results but fall short compared to CEDNN, particularly in capturing fine features and maintaining high segmentation quality.

The baseline for efficiency benchmarking was defined as a standard 3D U-Net implementation, trained and tested on the same dataset under an identical hardware configuration as CEDNN. This setup ensures that the comparison highlights architectural differences rather than variations in datasets or computing resources. Model size, training time, inference time, GPU memory usage, and FLOPs for both CEDNN and the 3D U-Net baseline are shown in Table 4, which highlights the computational advantage of the proposed method. Due to its widespread application in retinal OCT segmentation tasks, the 3D U-Net baseline [31] serves as a meaningful benchmark for evaluation.

The proposed approach performs significantly better than the current methods across a number of measures, shown by the computational efficiency analysis in Table 4, which makes it ideal for large data sets such as 101,000 images. The. The model achieves a 50% reduction in size and GPU memory usage, enabling deployment on resource-constrained devices. Training time is reduced by 58%, and inference time is halved, allowing for faster predictions, with the entire dataset processed in 18 minutes compared to 36 minutes with baseline methods. Additionally, the model's computational complexity is minimized, with a 50% reduction in FLOPs, ensuring quicker and more efficient operation without compromising accuracy. These improvements highlight the model's scalability, accessibility, and potential for real-time applications in clinical and computational settings.

Table 4. Computational efficiency analysis of CEDNN compared with a 3D U-Net baseline [31], trained and evaluated on the same dataset and hardware

The proposed method demonstrates strong performance in segmenting DR lesions, but like any approach, it does have certain limitations. Although the technique works well for typical DR symptoms, including blood vessels, microaneurysms, and exudates, it has difficulty discriminating tiny or subtle lesions because of low contrast or poor image resolution. Additionally, exudates, which often appear as bright spots, may sometimes be misclassified, especially in images with noise or artifacts. In scenarios where the image quality is compromised, such as low resolution or significant motion blur, the segmentation accuracy may decrease. To address these limitations, more advanced contextual information is needed to improve the segmentation of small or ambiguous lesions.

A novel approach to enhance retinal vascular segmentation and lesion detection in 3D-Spectral Optical Coherence Tomography (OCT) images is described in this work to assist with the early and precise diagnosis of diabetic retinopathy (DR). Local and global contextual information are efficiently extracted from retinal images using the Context Encoding Deep Neural Network (CEDNN) architecture, which is included into the proposed method. This allows for precise segmentation of retinal vessels and accurate identification of lesions indicative of DR progression. The proposed technique demonstrated excellent performance on a number of evaluation parameters, demonstrating its effectiveness in lesion identification and segmentation tasks. The experimental findings demonstrate the approach's resilience, with an exceptional 99.3% accuracy, 99.4% sensitivity, 99% specificity, 99.2% precision, and 99.3% F1-score. Furthermore, the predicted 95% Confidence Interval (CI) for accuracy falls between 98.8% and 99.8%. Diagnostic capacity is improved by the precise identification of lesions, such as microaneurysms, exudates, and hemorrhages, which enables early DR detection and immediate action. Because of its statistical resilience and high performance, the approach is a useful tool for clinical usage, especially in the early detection and treatment of DR. Despite the strong performance, the model’s near-perfect results raise concerns of possible overfitting and limited generalization to unseen populations. Future work will focus on validating the framework on independent external datasets and incorporating strategies such as domain adaptation and cross-center evaluation to ensure robustness across clinical settings.