GRIFD: a graded region-wise dissection and cross-pooling RNN framework for precise diabetic retinopathy detection in fundus images

RNN-based feature modeling

In the proposed work, a RNN is used as the core learning architecture to model temporal dependencies and spatial variations in DR lesions. The purpose of integrating the RNN is to iteratively process extracted feature sequences from fundus images and analyse changes in pixel intensity distribution across different ROI. The RNN’s memory state enables the system to retain contextual information from previously seen features, which is crucial for identifying subtle lesion progression or spatial inconsistencies that are not visible in isolated patches.

From the ROI examination, the maximum differences between features with higher variation are detected. Thus, the ROI-based examination is done for the maximum differences analysis. Hereafter, the following section focuses on the recurrent neural network where cross-pooling is observed to validate the variation process and improve the precision. The ROI is detected using maximum differentiated features in the sequential pixel distribution along the input image. The features with high variations within the identified ROI are dissected to determine the maximum changes. The following equation analyses the assessment layer used to store the previous data of the DR image.

(1) $$DR(n)+F_x=P_u\cdot R_0\cdot (n-1)+n^{\mathrm{\prime }}\cdot \sum _{v_a}\left(R_g+t_c\right).$$The assessment layer is processed to grasp the previous step of computation and produce the result. The assessment layer is responsible for performing the pixel identification from ROI and pursues the feature extraction process. This illustrates the feature extraction method determines the detection phase and it is pursuing the pixel distribution. The pixel distribution is done for variation identification and it is symbolized as $v_a$. Based on this assessment, layer RNN grasps the previous step from $n$ number of layers and examines the cross-pooling method, which is described as $R_g$. Post to this method, the memory state is used to reduce the complexity of DR image detection from the pixel distribution, and it is equated in the below derivative.

(2) $$m_a=R_0\left(n-1\right)\ast v_a+\left(g_h\ast F_x\right)+t_c.$$The memory state is $m_a$, the variation changes are observed and it is represented as $g_h$, this step uses the assessment layer under RNN and stores the information regarding DR. Based on the DR information the severity is analyzed reliably, by computing the variation process. The recurrent neural network model for variation detection is portrayed in Fig. 2.

Figure 3 shows how sequential features are passed through the RNN, how memory states preserve prior variation trends, and how cross-pooling detects dissimilarities between consecutive layers for accurate DR classification.

The RNN process mode for $v_a$ detection is illustrated in Fig. 2 for $1ton$ layers. The $R_o$ detected are the inputs for extracting $h_c$ using $F_x\mathrm{\forall }{o}_a$ and $b_r$. Using these variants of $F_x$, the $P_u$ range as 0 or 1 or $0\le ||<1$ is defined. The range of $P_u$ is normalized to identify if $F_x\in t_c$ is same $n^{\mathrm{\prime }}$ output $\in X^{\mathrm{\prime }}$. If the parameters are not the same, then $v_a$ is the identified factor for $n$ layer outputs. However, this variation is less concerned with maximum/minimum changes. The consecutive number of $h_c$ and $v_a$ changes (differences) are used to define $g_h$ provided the recurrency is maintained. The mediate $R_g$ impacted by the $v_a$ in any $n$ output is assessed for $y_i$ such that $n^{\mathrm{\prime }}$ is first performed. Irrespective of the $DR\left(0\right)$ to $DR\left(G\right)$ changes, the $i^{\mathrm{\prime }}$ is the least possible $v_a$ changes identified. If this $v_a$ change is detected then the $R_g$ changes are exactly detected. This variation is examined by using the cross-pooling method and defines the memory state that computes the storage of the previous set of information as the assessment layer does. Post to this method the cross-pooling is done to check whether the variation decreases or not and it is evaluated below.

(3) $$R_g=\prod _{m_a}^{P_u}\left(DR\ast v_a\right)+\left[\left(g_h\left(i^{\mathrm{\prime }}\right)\ast y^{\mathrm{\prime }}\right)\right]\ast DR(n).$$

The cross-pooling is performed to analyse the variation and whether it decreases. The minimum range is $i^{\mathrm{\prime }}$, where it validates the DR image and efficiently examines the cross-pooling. Thus, the cross-pooling is used in the memory state and acquires the assessment layer. The variation decreases and consistency is maintained, and these changes are used to observe the higher and lower range of variation detection. This format defines the changes, and from this violation, changes are observed in the following equation.

(4) $$d_m=m_a+\left(R_0+R_g\right)\ast \prod _{P_u}\left(x_0-i^{\mathrm{\prime }}\right)\ast g_h.$$

The determination is processed for the violation changes that occur due to the variation check and measures the maximum and minimum format. The maximum is represented as $x_0$, this stage uses the memory state for the identification of diseases and on which layer it relies. This determination is labelled as $d_m$, where it computes the maximum and minimum ranges and the detection phase. The pixel distribution is followed up to find the violation changes that have minimum changes in this step, training is given and it is equated as follows.

(5) $$t_r=d_m+m_a\ast \left(P_u\ast R_0\right)+DR\left(n\right)\ast V_d+y^{\mathrm{\prime }}-i_t.$$

The training is followed up in the above equation, which utilises the determination phase for analysing maximum and minimum values. This relies on the cross-pooling method and finds the variation detection and from this violation, changes are detected. If the violation changes are recognised, then the training is given and it is labelled as $t_r$.

The selection of a RNN to be employed in the GRIFD framework is intentional and driven by its strength in modelling sequential dependencies in pixel-level feature changes between ROI blocks. In contrast to CNNs, which focus only on local spatial patterns, and Transformer models, which utilise global self-attention but are computationally expensive, the RNN efficiently models the evolution of features in spatially contiguous areas with minimal overhead. This works well in fundus images where lesions frequently appear as gradual, sequential changes and are less likely to appear as sudden changes. The memory state of the RNN allows the model to follow these fine changes across time and support the identification of accurate lesion boundaries while inhibiting inconsistent or false positive areas. Such sequential modelling, when added with our cross-pooling mechanism, is the strongest point of GRIFD in supporting higher precision and specificity along with minimal detection delay.

Proposed GRIFD framework

In the envisioned GRIFD model, the RNN module utilises a memory cell $c_t$ and a hidden state $h_t$ to track sequential shifts in feature activations between consecutive ROI blocks. This is borrowed from the temporal modelling properties of LSTM-based networks, commonly used in longitudinal medical image analysis. The cross-pooling violation pertains to situations where neighbouring regions produce contradictory pooled features, which are probably due to image noise or intra-anatomical structural overlap. To counter this, our cross-pooling procedure corrects inter-region continuity in accordance with ideas utilized in spatial feature calibration observed in attention-guided filtering approaches.

The feature extraction is done from the input DR image, and the features associated with the three stages are identified: mild, regular, and high. This examination step illustrates the variation process that is checked periodically under ROI. This extraction enables the accurate detection of DR, which is crucial for defining the recurrent neural network. This network represents the region in n x m matrices and finds its features. Here, the characteristics of the image are analysed, and the changes are exhibited under the variation process. This step defines the pixel distribution from the ROI and estimates the higher-level feature extraction process. The desired features are extracted reliably, and training is provided if any violations occur. Due to these violations, the constant range is not maintained under RNN. From this part of the study, the precision degrades, necessitating improvement in this examination step. The initial step is to validate the DR image using a normalisation approach, which is described below.

(6) $$V_d=DR+F_x\cdot (b_r+o_a)+\frac{1}{DR}\cdot \prod _{b_r}^{o_a}\left(F_x+{\mathrm{\Delta }}_t\right).$$

(7) $${\mathrm{\Delta }}_t=\left((DR+F_x)\cdot (b_r+o_a)\right)+y^{\mathrm{\prime }}-i_t+\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{z}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}(F_x,DR).$$

The validation is carried out in the above equation and it is represented as $V_d$, the DR input image is symbolised as DR, and the feature extraction is denoted as $F_x$, the brightness and contrast are termed as $b_r$ and $o_a$. The periodic monitoring is represented as $y^{\mathrm{\prime }}$, $i_t$ is the time taken to perform the particular processing step. The evaluation occurs for the different feature extraction processes from the input DR image. Here, periodic monitoring is observed better, and the extraction is pursued reliably. Both the brightness and contrast are considered, and better analysis is performed for image retrieval from the database and detecting whether the diseases are standard, higher, or lower. Executing this step determines the period that considers the image’s brightness and contrast and performs better detection to improve the precision level, which is the scope of this work. The progression is executed reliably and illustrates the pixel-distribution stage.

Pixel distribution assessment

Before training the GRIFD model, two types of inputs were extracted from the fundus images: Feature-Based Inputs, these include textural and intensity-based descriptors such as mean intensity, local binary patterns (LBP), and histogram-based contrast variations computed within each ROI block. Each image was partitioned into non-overlapping n $\times$ n blocks, from which first-order statistics and Gabor-filtered responses were collected as features. Edge-based inputs, for edge-oriented analysis, the same ROI blocks were passed through gradient-based filters (e.g., Sobel and Laplacian) to extract contour and vessel boundary features. Edge maps were then analysed for abrupt variation transitions using variance and entropy calculations across neighbouring pixel clusters. These extracted feature vectors were then standardised and normalised before being used as input to the recurrent neural network. The varying conditions (i.e., different lighting, contrast, lesion presence, and vessel occlusion) were simulated across images to evaluate the model’s robustness under multiple feature and edge conditions.

The pixel distribution assessment is performed using the extracted features; the initial stage is to validate the DR image using normalization. This normalization is carried out based on the pixel-distribution process, which defines the validation method. The validation is executed for the DR image extraction from the database and performs the normalisation that includes the brightness and contrast. Based on these two parameters, the feature of the DR is extracted and observes the better detection of variation. Here, the variation check is executed to examine the reliable processing step and verify the normalisation condition. This normalization is used to illustrate whether the image has higher brightness or contrast, and in other cases, lower brightness or contrast. On processing this evaluation, the validation runs through reliable feature extraction promptly. From this computation step, the analysis is conducted to detect DR within the specified severity range, and it identifies the specific characteristic, as formulated below.

(8) $$y_i=(DR+F_x)\cdot \sum _{b_r}^{c_a}\left(h_c\cdot h_c+\gamma \right).$$

(9) $$\gamma =⟨(P_u+F_x)\cdot t_c⟩+\left[(t_c\cdot P_u)\cdot (F_x+h_c)\right]+\sum _{t_c}\left((h_c+DR)\cdot [y^{\mathrm{\prime }}-i_t]\right)+\mathrm{R}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}.$$

The analysis is conducted for the specific characteristic used to detect DR images. The characteristic is represented as $h_c$, $P_u$ is symbolised as pixel-distribution, the detection is $t_c$. This is executed in the desired manner and that illustrates the periodic observation takes place on time. The evaluation step is executed for the varying DR image and is associated with the characteristic used for the analysis phase. This analysis, denoted as $y_i$, is performed under different input image sets and examines pixel-based acquisition. This defines a more efficient processing step for the DR image and facilitates the variation check. The variation check is followed up for the DR and analysis of the characteristic, which refers to the light condition during image capture. Based on this capturing process, the study shows a slight variation in the input image. To improve the detection phase, the feature used to examine the characteristic-based input image is evaluated. This shows the image characteristic in the required manner for time-based computation.

Periodic observation is conducted to identify characteristic image features and accurately validate the pixel distribution. This evaluation is undertaken to facilitate the desired computation related to the timely observation of characteristics and provides the necessary steps accordingly. In this execution step, the analysis for the specific characteristic leads to detecting pixels on n × m matrices and gives the necessary step. This processing step is used to identify the DR image along with the features and characteristics and provides the appropriate detection phase. This includes the appropriate processing that estimates the memory state, which involves mapping the DR image’s previous data to the current data and is executed under the RNN. The computation performs feature extraction to provide necessary detection for characteristics and gives the parameters. This processing step is carried out based on the characteristics and features of the DR image. Following this observation process, feature extraction is performed using the equation below.

(10) $$F_x=y^{\mathrm{\prime }}+(DR\cdot t_c)+\sum _{P_u}^{V_d}\left((y_i+DR)\cdot \left\{\left[(P_u+R_0)+(DR\cdot h_c)\right]\cdot (t_c-i_t)\right\}\right)+\mathrm{R}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}.$$

(11) $$\mathrm{R}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{d}\mathrm{u}\mathrm{a}\mathrm{l}=\sum \left((DR\cdot V_d)+\left(n^{\mathrm{\prime }}(h_c)\cdot V_d\right)-i_t+y^{\mathrm{\prime }}\cdot y_i\cdot (h_c+DR)\right).$$

The feature extraction is executed in the above equation, which represents the DR image and is associated with a characteristic of the image. The evaluation takes place for the normalisation of DR, as discussed in equation, where the ROI is symbolised as $R_0$. The normalisation is described as $n^{\mathrm{\prime }}$, the processing step is used to determine a better detection of pixel distribution for the improvement of precision. The precision level is maintained reliably, and it estimates the image characteristics process. The progression is used to define the image characteristic and find the validation and normalization of the input image. This input image is used to extract the necessary features and perform the detection phase. The detection plays a vital role in the validation process, exhibits the image’s feature and computes the variation. The pixel distribution is performed using DR detection and displays the ROI resulting from the pixel distribution. The $R_0$ selection process is defined using Fig. 3.

The Fig. 3 depicts how the proposed method selects ROI blocks by intersecting pixel-distributed features and evaluating them against feature variation thresholds to determine eligible dissection zones.

The $R_0$ is identified as an intersection between $\left(m\times n\right)\mathrm{\forall }{y}_i$ and $P_u$ extracted from $i_t$. Depending on the $F_x$ identified, if $\left(h_c\le F_x\right)$ then $o_a$ and $b_r$ for $n$ and $m$ are used to define the new region of $D{R}_i\notin P_u$. If the range exceeds the actual $P_u$ then $n^{\mathrm{\prime }}$ is pursued as $\left(DR\ast V_d\right)$ and $\left(y^{\mathrm{\prime }}-i_t\right)$ for all $\left(P_u+F_x\right)\ne 0$. This process is computed towards $b_r$ and $o_a$ classification for $h_c$ differentiation such that $n^{\mathrm{\prime }}\left(h_c\right)$ is the sub-category of normalization performed from $\left(P_u+R_o\right)$ whereas, $\left(y^{\mathrm{\prime }}+h_c\right)$. Based on this difference between $y^{\mathrm{\prime }}$ and $P_u$, the $R_o$ is extracted. Hence, the process constitutes both $o_a$ and $b_r\in h_c$ (Fig. 3). The extraction is done for the varying DR that is executed in the required manner and illustrates the $n\times m$ matrices. In this methodology, validation is exhibited for the DR normalization process and evaluates the precision rate. By examining the severity range, the diseases are detected with the use of features and characteristics and pursue the extraction of necessary features. Only the necessary features are extracted within the required period, and a more thorough analysis of the memory state enables the periodic checking of variations that occur under ROI. Hereafter, the feature extraction is an important step that calculates the ROI using the DR image as input and exhibits validation. Validation and normalisation are used for the feature extraction, which includes the brightness and contrast. Based on this, the ROI is calculated using the pixel distribution method. From this approach, the necessary features are extracted from the DR, and a post-processing step is applied to the pixel distribution based on the extracted features, as formulated below.

(12) $$P_u=DR\cdot (t_c+V_d)\cdot R_0+(b_r+o_a)+(F_x+n^{\mathrm{\prime }})\cdot V_d+F_x\cdot DR+(F_x\cdot i_t)\cdot y_i.$$

The pixel distribution is done from the feature extraction process and that defines the better recognition of brightness and contrast. To identify the necessary features the pixel states the collection of blocks that appear on the DR image. From the feature, the segmentation is done as a pixel distribution that includes the block size. Here, the n × m is used to define the feature extraction process and periodically pursues the checking process. The n × m matrices are used under pixel distribution and they provide the appropriate detection of DR image. This processing step is used to define the matrix format in which the feature extraction is performed. The matrices used in this work are used to matrix the feature extraction process. The pixel distribution is carried out in the desired format, which exhibits the feature for the change area derived from the pixel-based distribution obtained through detection in the region. The ROI is used to determine this pixel distribution and provides the necessary steps accordingly. The brightness and contrast are used to evaluate the differences among the features.

The feature extraction is evaluated for the pixel distribution process and that defines the validation process for $n\times m$ matrices. The ROI is calculated for the varying analyses and findings related to the severity of diseases at different stages. The normalization is followed up for the specific characteristic and estimates the desired processing. The progression runs through the analysis of the maximum difference between the features and with a higher variation rate. This variation rate differs based on the pixel-distribution technique, which exhibits the necessary feature extraction. The essential features extracted along with these characteristics are used to examine the periodic validation of DR and find the treatment. The treatment is carried out for the DR patient by identifying the region of interest, which is performed under pixel distribution. Hereafter, the examination is done for the ROI where the maximum difference between the features with higher variation is computed and expressed in the equation below.

(13) $$X^{\mathrm{\prime }}=R_0+\left(F_x\ast P_u\right)+y_i\ast DR+y^{\mathrm{\prime }}.$$

The examination is done for the ROI features and it is described as X′, this includes pixel distribution and feature extraction. The processing step includes the n × m matrices that pursue the DR image and analysis is carried out. The analysis is followed up for the feature extraction and pixel distribution process, and defines a better recognition of DR from the input image. The examination is done for the ROI image that exhibits the detection phase and it is associated with the pixel distribution. From this, both the necessary features and pixel distribution are followed up on periodically to detect the DR.

Cross-pooling mechanism

From this estimation, variation detection is followed up as the iteration from RNN for change detection. The variations are the training inputs used for the detection of lower variation occurrences. Hereafter, the variation detection is examined in the following equations.

(14) $$v_a\left(g_h\right)=y^{\mathrm{\prime }}+\left(DR\left(n\right)\ast F_x\right)+v_h-t_r.$$

The variation change occurrences are analysed in the above equation, which exhibits the training phase for the minimum variation. These changes are done from the violation step, and it is labelled as $v_h$, and evaluates the recognition of variation from high to low. If it occurs, then the change detection is carried out for the variation process, where the classification is processed, and it is represented in the equation below as follows.

(15) $$g_h\left(t_c\right)=v_a\left(x_0,i^{\mathrm{\prime }}\right)\ast P_u+R_0\left(DR\right)\ast f_s.$$

The change detection is performed and finds the ranges from high to low variant, from this case, the pixel distribution and feature extraction verifies the change detection occurrences reliably. The classification $f_y$, is used to differentiate the maximum and minimum variation from the cross-pooling process among the layer’s representation. The cross-pooling process variation detection is portrayed in Fig. 4.

Figure 4 presents the evaluation of variation levels (high or low) across RNN layers through cross-pooling, enabling the network to classify regions with precise control over changes and dissection logic.

The cross-pooling for $v_a\left(g_h\right)$ detection relies on $v_a$ observed in $i_t$ and $n\mathrm{\forall }{P}_u$. Based on the $\left(i^{\mathrm{\prime }}\times y^{\mathrm{\prime }}\right)$ input classification for $x_o$ and $\left(x_o-i^{\mathrm{\prime }}\right)$ differentiations, the $v_h$ check is performed. From this computation, $\left(t_c-i_t\right)$ and $\left(t_c+i_t\right)$ for any range of $P_u$. Considering this validation, the $DR(0)$ to $DR\left(n\right)$ is clubbed from the last known $m_a$ and $i^{\mathrm{\prime }}$ detected. If the $i^{\mathrm{\prime }}$ is less than the previous $P_u$ instance, then $\left(x_o+i\right)$ is the maximum change identified from $g_h$. In case of a new $v_a$ detected from any $n$, the $R_g$ is estimated from $t_r$ provided $\left(v_h-t_r\right)$ is the recent (updated) change. Therefore, the change is suppressed by supplementing $b_r$ (or) $o_a$ from the $h_c$ detected. Therefore, $v_a\left(g_h\right)$ classifies $\left(i^{\mathrm{\prime }}\times m\right)$ and $\left(i^{\mathrm{\prime }}\times n\right)\mathrm{\forall }{v}_h$ check as in Fig. 4. Thus, the change detection is executed for the classification model, and from this region, dissection is done from the change analyzed as maximum to minimum and it is identified below.

(16) $$f_y=g_h\left(t_c\right)+t_r\ast V_d\left(F_x+t_c\right)\ast y^{\mathrm{\prime }}.$$

The identification is observed for the analysis of the maximum to minimum range of variation process and that illustrates the validation method for the feature extraction. The identification is termed as $f_y$, which detects the phase for DR images and determines the cross-pooling techniques. The detection is carried out for the region dissected from the occurrences of the change in the variation process. The precision level improves, as shown below.

(17) $$\alpha =f_y+\left(DR\ast t_c\right)+\left(x_0,i^{\mathrm{\prime }}\right)\ast V_d+R_g(h_c).$$

The precision level improves by decreasing the error rate, which is achieved through this reduction, as discussed in Eq. (2). The precision is $\alpha$, which defines the validation process for the cross-pooling method and provides the change occurrences from maximum to minimum variation; thus, region dissection is done using RNN. Therefore, the scope of this article is satisfied by using RNN for region dissection.

Hyperparameter analysis

In the hyperparameter analysis, the variables related to dissection accuracy are identified and their impact on performance is studied. In this manner, the $n^{\mathrm{\prime }}$ confusion matrix for $X^{\mathrm{\prime }}\mathrm{\forall }{g}_h$ and $v_a$ is represented in Fig. 5.

The confusion matrix shows the classification performance across different classes of DR severity, highlighting true positives, false positives, and false negatives obtained from the GRIFD.

The $v_a$ and $g_h$ changes are abrupt by identifying $\left(x_a+i^{\mathrm{\prime }}\right),\left(x_o\right),and\left(x_o-i^{\mathrm{\prime }}\right)$ from different $i_t$. The $t_r$ for $DR\left(0\right)$ to $DR\left(n\right)$ is performed to verify if $\left(h_c\le F_x\right)$ for any range of $\left(m\times n\right)$. The categorization of $\left(i^{\mathrm{\prime }}\times m\right)$ and $\left(i^{\mathrm{\prime }}\times n\right)$ are performed under the above three variants of $x_o$ such that $R_g$ delivers errorless $v_a$. The recurrent processes of $\left(t_c-i_t\right)$ and $\left(t_c+i_t\right)$ is required to decide $n^{\mathrm{\prime }}\mathrm{\forall }{h}_c$. Hence, the $y_i$ for the varying sizes is optimal to reduce the $v_a\left(g_h\right)$. The $R_g$ identifies the $i^{\mathrm{\prime }}$ from the consecutive set of $F_x$ and $h_c$ mapping to ensure fewer errors are identified. Thus, the new $t_r$ relies on the target of $\left(x_o-i\right)$ for reducing the $v_h$ that maximizes the dissection of $\left(i^{\mathrm{\prime }}\times m\right)$ and $\left(i^{\mathrm{\prime }}\times n\right)$. Therefore, the $n^{\mathrm{\prime }}$ outputs are sufficient to maximize the dissection accuracy (Fig. 5). The ROC for $v_a$ and $g_h$ is analysedmaximise as presented in Fig. 6.

The receiver operating characteristic curve illustrates the TPR against the FPR for different thresholds, demonstrating the classification capability and sensitivity of the proposed method.

The true positive rate (TPR) for the false positive rate (FPR) for $v_a$ and $g_h$ is analyzed in Fig. 6. The $\left(x_o-i^{\mathrm{\prime }}\right)$ is the objective for accuracy improvement regardless of $x_o$ and $\left(t_c+i_t\right)$. The recurrent learning process identifies the exact $\left(P_u-n^{\mathrm{\prime }}\right)$ and $\left(P_u+n^{\mathrm{\prime }}\right)$ variants from the actual $F_x$. In the condition validation of $\left(F_x\ge h_c\right)$, the $R_g$ with maximum $\left(x_o\right)$ is suppressed. This reduces the chance of FPR for $\left(m\times n\right)$ by performing $\left(i^{\mathrm{\prime }}\times m\right)$ and $\left(i^{\mathrm{\prime }}\times n\right)f_y$ such that existing $F_x$ is sufficient for increasing accuracy. The $\left(DR\ast t_c\right)$ and $R_g\left(h_c\right)$ regulates the $x_o$ by identifying $d_m$ from $m_a$ to ensure $\left[g_h\left(i^{\mathrm{\prime }}\right)\ast y^{\mathrm{\prime }}\right]$ reduces the error and thereby the training retains the maximum ${\mathrm{\Pi }}_{P_u}\left(x_o-i^{\mathrm{\prime }}\right)$ for accuracy improvement. In the final hyperparameter analysis, the error is considered and presented in Fig. 7.

Figure 7 illustrates the convergence pattern of the error metric over multiple training iterations, indicating the robustness of the model for different feature and edge types.

The error analysis for the $t_r$ iterations and $n^{\mathrm{\prime }}$ is presented in the above Fig. 7. The $\left(P_u+F_x\right)\ne 0$ is the optimal case for reducing $v_a$ and $g_h$ under different $i_t$. Based on the $DR\left(0\right)$ to $DR\left(n\right)$ outputs the $n^{\mathrm{\prime }}$ priority is defined; this instance identifies the $x_o$ to $\left(x_o+i^{\mathrm{\prime }}\right)$ or $\left(x_o-i^{\mathrm{\prime }}\right)$ variant by identify $R_g$ changes. Depending on the $o_a$ and $b_r$ differences in $t_c$ from any $n$ layers, the $\alpha$ improvement is pursued. Therefore, the $R_g$ is revisited using $m_a$ that validates $d_m$ to reduce $v_h$. If this is achieved then X′ for all $P_u$ ad $F_x$ are equated for similar congruency and specificity. Therefore, the least $i^{\mathrm{\prime }}$ serves as the $t_r$ point for any range of FPR observed. The region combining $m$ and $i$ is segregated from $n$ to $i$ to ensure fewer errors under $t_r$ and $n^{\mathrm{\prime }}$.

Experimental setup and results

The proposed method is evaluated using MATLAB codes executed in a standalone computer with 4 GB-2slots random memory and a 2.1 GHz Intel processor with a storage of 128 GB.

Before training the GRIFD model, two types of inputs were extracted from the fundus images: Feature-based inputs include textural and intensity-based descriptors, such as mean intensity, local binary patterns (LBP), and histogram-based contrast variations computed within each ROI block. Each image was partitioned into non-overlapping n $\times$ n blocks, from which first-order statistics and Gabor-filtered responses were collected as features. For edge-oriented analysis, the same ROI blocks were passed through gradient-based filters (e.g., Sobel and Laplacian) to extract contour and vessel boundary features. Edge maps were then analysed for abrupt variation transitions using variance and entropy calculations across neighbouring pixel clusters. These extracted feature vectors were then standardised and normalised before being used as input to the recurrent neural network. The varying conditions (i.e., different lighting, contrast, lesion presence, and vessel occlusion) were simulated across images to evaluate the model’s robustness under multiple feature and edge conditions.

Dataset

The testing fundus image inputs are fetched from Akram et al. (2020) which contains 100+ fundus images of patients between 25 and 80 years of age. The digital image resolution is $1\text{,}500\times 1\text{,}000$ pixels of which 76 are infected with DR. The learning rate is set between 0.8 and 1.0 for 1,200 iterations under three epochs. The training images are 100 and the testing image count is 76 for which the epochs are repeated for maximum precision. This dataset contains retinal blood vessel networks, segmented artery and vein networks, and a testing image count of 76. It is used to calculate the arteriovenous ratio (AVR), annotate the optic nerve head (ONH), and identify various retinal abnormalities, such as hard exudates (HE) and cotton wool spots. we employed a grid search strategy combined with 5-fold cross-validation on the Mendeley dataset to identify optimal values for learning rate, batch size, number of epochs, and dropout rate. Performance was evaluated based on validation accuracy and mean error across folds.

These images are organised into distinct categories based on retinal conditions, including DR, hypertensive retinopathy, papilledema, and normal. While the dataset provides folder-level grouping for these four diagnostic classes, it does not include explicit annotations for the different stages of DR (such as mild, moderate, or severe). Therefore, any further categorisation by DR severity must be performed manually based on visual inspection of retinal lesions such as microaneurysms, haemorrhages, and exudates, or through clinical labelling where available. This dataset serves as a valuable resource for evaluating both segmentation and classification algorithms in the context of retinal disease detection. Figure 8 presents the identified ROIs from test images using the GRIFD model, with a focus on severity-based visual segmentation across different DR stages.

Figure 9 presents visual or matrix-based results of lesion region segmentation and dissection after applying feature variation analysis and RNN-based validation.

The dataset for the current study comprises more than 100 colour fundus images from the Mendeley Data repository (Akram et al., 2020), which contain various retinal conditions, including diabetic retinopathy. A total of 100 images were manually examined and selected for training and testing, 24 normal, 28 mild DR and 48 severe DR—a balance across the spectrum of severity. All images were rescaled to 512 $\times$ 512 and underwent preprocessing operations, including grayscale conversion, contrast normalisation, histogram equalisation, and median filtering for denoising. ROI-based cropping was used to focus on central retinal areas. Annotation validity was ensured by referencing the metadata and folder structure in the dataset. In instances where severity labels were not available, clinical features such as microaneurysms and haemorrhages were visually examined and labelled by expert annotators to preserve the ground truth.

A dropout layer with a rate of 0.3 was applied after the RNN and fully connected layers to reduce co-adaptation of neurons and minimise overfitting. Additionally, L2 regularization with a coefficient of 1e−4 was applied to the weights during training. For the Mendeley dataset, we applied basic augmentation techniques including horizontal flipping, rotation ( $\pm$15 degrees), brightness scaling ( $\pm$10%), and minor Gaussian blurring to synthetically expand the training set and introduce variability in lesion appearance. These augmentations were applied in real time during each training epoch.

Comparative analysis

The comparative analysis is presented using the metrics: accuracy, precision, specificity, mean error, and segment detection time. These metrics are analysed under the varying features and edges, and therefore, they vary in number (1–12) and (1–9), respectively. To verify the proposed method’s efficiency, the method is compared with the existing RTNet (Huang et al., 2022), MCNN-UNet (Skouta et al., 2022), and DRFEC (Das, Biswas & Bandyopadhyay, 2023) methods.

To assess the performance of the proposed GRIFD method, the following standard metrics are used: Accuracy (Acc) is the proportion of correctly predicted observations (both DR and non-DR) out of the total observations.

(18) $$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\frac{TP+TN}{TP+TN+FP+FN}.$$

Precision (Prec) is the ratio of true positive predictions to the total predicted positive cases. It reflects the model’s ability to correctly identify DR.

(19) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{TP}{TP+FP}.$$

Specificity (Spec) is the ability of the model to correctly identify non-DR cases (true negatives).

(20) $$\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}=\frac{TN}{TN+FP}.$$

Mean error (ME), is the average absolute error between predicted and actual region boundaries or lesion locations. This is computed across all tested images.

(21) $$\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{r}\mathrm{o}\mathrm{r}=\frac{1}{N}\sum _{i=1}^N\left|P_i-A_i\right|$$where $P_i$ is the predicted value and $A_i$ is the actual ground truth. Detection delay ( $D_d$) is the average computational time (in seconds) taken by the model to detect and localise the DR region per image.

To conduct fair and consistent comparison all three methods (Huang et al., 2022; Skouta et al., 2022 and Das, Biswas & Bandyopadhyay, 2023) were re-implemented and trained on the same fundus image dataset used in the proposed GRIDF method. RTNet is a transformer-based architecture that applies self-attention mechanisms to establish relational dependencies between lesions and their corresponding vessel features in fundus images. It combines the accuracy of a relation encoder-decoder framework to perform multi-lesion segmentation. The MCNN-UNet method combines a multi-channel CNN feature extractor with a U-Net-inspired decoder for the task of hemorrhagic regions segmentation, preserving the channel-wise feature hierarchies. The DRFEC method consists of a convolutional feature extraction process and a fully connected classification process to identify and compare severity of DR-related abnormalities due to their structure and texture in associated vessels. The number of images, training methods and parameters, and the dimensions of what the network sees remained consistent and each method was trained from scratch. For both training (100 images) and testing (76 images), the same image set from the Mendeley Data (Akram et al., 2020) repository was used. These scenarios enable the evaluation of the most expansive metrics, including accuracy, precision, specificity, mean error, and detection delay,the through both training and testing across the three methods.

Accuracy

Figure 10 compares the accuracy metric for the proposed and existing methods across multiple feature and edge variations, showing superior performance of GRIFD.

The accuracy rate for the proposed work is high (Fig. 10) for varying features and edges and that purses the feature extraction process and estimates the pixel distribution. The input DR image is acquired, and the feature extraction process is performed here, resulting in n × m matrices that are accurately followed up. This illustrates the better recognition of ROI from the two matrices and provides the distribution factor. The validation is performed for image normalisation, which in turn supports the severity analysis for disease detection. The varying stages are used to explore the ROI and provide an estimation process by using an RNN. The necessary feature is extracted by detecting changes in occurrence. Here, the training is given to the RNN and that is associated with the cross-pooling process. The accuracy rate for the proposed work is used to determine the pixel distribution, based on the extraction. Achieving accuracy plays a significant role in this proposed work and employs the normalization of the DR image. This is associated with the severity caused by the diseases and address and it is formulated as $\left[\left(DR+F_x\right)\ast \left(b_r+o_a\right)\right]+y^{\mathrm{\prime }}$.

Precision

Figure 11 shows how GRIFD outperforms other methods in terms of precision by effectively identifying regions with maximum variation across fundus images.

The precision in this work is improved (Fig. 11) by identifying the maximum difference between the features that show the higher variation changes. This is observed in two metrics such as features and edges of the DR image and that is associated with the pixel distribution. Based on this proposal the validation is followed by the training phase that uses the memory state under RNN. This methodology illustrates the feature extraction process and defines the specific characteristics of the image. Both the image’s feature and unique characteristics are considered to give the necessary result. In this manner, it determines the accuracy of the DR image that envelopes the examination step and provides the n × m matrices. This processing step defines the normalization of the image by validation process and examines the violation changes. The violation changes occur due to the feature extraction process and accurately define the precision level and it is equated as $⟨\left(P_u+F_x\right)\ast t_c⟩+\left[\left(t_c\ast P_u\right)\ast \left(F_x+h_c\right)\right]$, here it shows a higher precision rate.

Specificity

The Fig. 12 compares the specificity scores, reflecting how accurately each method avoids false positives while detecting DR-affected regions.

The specificity is computed for the pixel distribution and uses $m_a$, for the DR images. This estimates the necessary feature extraction process that illustrates the variation change occurrences. The variation changes define the reliable specificity among the DR images and detect the diseases. Periodic monitoring is observed in this category, which determines the precision level of accuracy. This methodology is developed to explore the variation and gives the feature-based processing for the features. These features are associated with the variation changes and provide the training phase. The training is given based on this violation changes are analyzed based on the pixel-based image. From this processing step, the extraction is monitored for n × m matrices, and the resultant is generated accurately. This section focuses on two varying metrics such as features and edges; based on these the specificity shows a higher value range. To explore the differences between the features that rely on higher variation and determine the specificity it is represented as $\left\{\left[\left(P_u+R_0\right)+\left(DR\ast h_c\right)\right]\ast \left(t_c-i_t\right)\right\}$ (Fig. 12).

Mean error

Figure 13 shows the mean error of region detection for all compared methods, indicating significant improvements with the proposed cross-pooling-based dissection.

In Fig. 13, the mean error is found to be less based on the specific characteristics and features. Based on this approach, the training phase is carried out in the desired manner, resulting in matrix-oriented processing. This matrix shows the pixel distribution among the input image and that is forwarded to the ROI. The ROI image is extracted from the pixel distribution, which relates to the cross-pooling mechanism. The pooling is used to check for the variation, ranging from higher to lower. If it is lower, the necessary training is given based on the feature distribution. The distribution is followed up for the violation changes and detects for the region distributed process. The scope of this work is to attain the region disserted and observe the variation process among the features. The cross-pooling is carried out in the desired way that explores the addressing of mean error in this work and reduces it is equated in Eq. (17) where the precision and error rate are equated and it is symbolised as $f_y+\left(DR\ast t_c\right)+\left(x_0,i^{\mathrm{\prime }}\right)\ast V_d$.

Detection delay

Figure 14 highlights the efficiency of GRIFD in reducing detection time, making it more suitable for real-time or rapid screening scenarios.

The detection delay is reduced in this proposed work which shows lesser value compared to the previous methods (Fig. 14). This progresses under two metrics such as features and edges, and gives the ROI-based image processing. The image processing runs through the desired computation and that relates with the region dissected. From this the precision level is improved and the error rate is reduced by considering this process so that the delay is addressed and shows a better DR image. The DR image is used to envelop the cross-pooling mechanism and relates with the pixel distribution among n × m matrices. Both the matrices are associated with changes occurrences due to the variation that is performed under variation changes. Thus, it explores the violation changes given to the training section and examines the better pixel distribution phase. This step is processed among the specific characteristics to explore the variation changes and reduce the delay factor. This detection is associated with the ROI region extraction and delay is addressed and it is formulated as $g_h\left(t_c\right)+t_r\ast V_d\left(F_x+t_c\right)\ast y^{\mathrm{\prime }}$.

Comparison with literature survey

A few recent works suggested varied methods to enhance diabetic retinopathy (DR) detection with their respective accuracy rates and architectural designs. Huang et al. (2022) presented RTNet as a model utilising transformers, along with relational self-attention, for multi-lesion segmentation, achieving an accuracy of 92.2%. Although RTNet excels at structurally modelling the dependencies of features across spatial context spaces, the model does not support sequential dissection, thereby precluding the ability to extract the subtle progression of the lesion over time. Mishra, Pandey & Singh (2024) employed an ensemble deep neural network (DNN) framework, achieving 88% accuracy. Though ensemble methods often enhance robustness, their performance may degrade in the presence of noisy or overlapping features due to a lack of explicit region-level dissection.

Guo & Peng (2022) put forwarded the model of CARNet, where attentive refinement was used for DR segmentation with multi-lesion and attained good accuracy of 93%. However, the model of CARNet becoming space-attention-fusion-dependent and lacking temporal context verification sometimes causes inconsistency in the identification of lesion boundaries at different pixel intensities. Wong, Juwono & Apriono (2023) used a transfer learning approach with parameter optimization and ECOC ensemble classification, but reported only 82.1% accuracy. This lower performance may be attributed to domain mismatch and limited adaptability of pre-trained features to the diverse lesion types in fundus images. Singh & Dobhal (2024) reported 92.66% accuracy using a DL-based transfer learning model, demonstrating competitive results. However, the model’s dependence on static feature representations constrains its ability to model evolving lesion characteristics across image sequences. Table 1 provides a comparison of results with the literature survey.

Singh, Gupta & Dung (2024) presented a fine-tuned deep learning model, but achieved only 74.58% accuracy, indicating a lack of generalization capacity, possibly due to limited training diversity or insufficient feature dissection. Naz et al. (2024) presented an enhanced fuzzy local information k-means clustering approach, which provided the best accuracy of 94.00% compared to the methods discussed above. Though encouraging, the clustering-based scheme has limited contextual learning over depth and is prone to hyperparameter sensitivity when the distribution of the lesion is highly variable. With the above methods suggested by each author in the above table as comparison methods, the suggested GRIFD method had accuracy rates of 94.35% (features) and 93.72% (edges). This work rivals each of the above methods or outperforms them. The edge of GRIFD lies in its graded region-of-interest dissection, cross-pooling-based variation filtering and RNN-based sequential learning simultaneously allow the model to actively abolish false positives and stably retain lesion structures throughout the ROI. Unlike static methods based on attention alone or transfer learning alone, GRIFD formally verifies the pixel-level consistency and derives the high-variation regions by itself and hence ensures precise segmentation even under challenging image conditions.

Cross validation and comparison with additional dataset

To further test the generalizability of the suggested GRIFD approach, we experimented with the publicly accessible EyePACS dataset, a large set of diverse colour fundus images with diabetic retinopathy severity level annotations. We randomly selected 1,000 images (balanced by normal, mild, moderate, and severe diabetic retinopathy) from the EyePACS dataset for testing. These were preprocessed using the identical pipeline outlined above. The GRIFD model was also trained employing a 5-fold cross-validation approach to guarantee robustness. Results showed uniform performance across folds, with a mean accuracy of 93.85%, precision of 94.20%, and specificity of 92.78%, validating the method on a larger dataset with enhanced variability in image quality and lesion distribution, as shown in Table 2.

To evaluate the generalizability and robustness of the presented GRIFD model, we conducted 5-fold cross-validation on the original Mendeley dataset as well as on the larger EyePACS dataset. As presented in Table 3, GRIFD performed uniformly well for all folds on the EyePACS dataset with an average accuracy of 93.85%, precision of 94.20%, and specificity of 92.78%. The detection delay remained constant at 1.00 s, proving the suitability of the method for real-world application. This performance ensures that the GRIFD algorithm maintains accuracy on a larger and more varied dataset, where differences in image quality, lesion size, and class distribution are greater.

Table 3 shows the results of 5-fold cross-validation on the Mendeley dataset. GRIFD had excellent accuracy for all folds, having an average accuracy of 94.16%, precision of 96.05%, and specificity of 94.83%, and a smaller mean error of 0.0544. The mean detection delay was 0.958 s, which reflects the model’s low-latency nature. These findings confirm that GRIFD is not only accurate in its predictions but also consistent across diverse subsets of the same set.

A comparative overview in Table 4 identifies that GRIFD was slightly better on the Mendeley dataset (in single-run as well as cross-validation environments) than on EyePACS, as could be anticipated from the fact that Mendeley contains tidier, more uniformly preprocessed images. Yet the slight degression in EyePACS performance reaffirms the model’s ability to generalize, as it continues to post robust metrics without significant loss of performance. This shows that the suggested GRIFD architecture can transfer successfully to other datasets with maintaining both diagnostic accuracy and computational cost.

Ablation study

An extensive ablation analysis was performed to analyse the effect of every main component in the GRIFD framework. The complete model achieved the highest performance overall, with 94.16% accuracy, 96.05% precision, and the most minor mean error and detection delay. The elimination of the cross-pooling layer resulted in a significant decline in specificity and precision, underscoring its crucial role in filtering out unreliable feature variations. Removing the RNN module impaired sequential learning, bringing accuracy down to 89.45%, replacing graded ROI dissection with uniform segmentation affected region localization and boosted the false positive rate. Last but not least, skipping preprocessing improvements yielded the worst accuracy (88.90%) and error rate due to unnormalized contrast and brightness levels in the input images. These findings indicate clearly that both modules are making significant contributions to the system overall, with cross-pooling and RNN taking centre stage in enhancing lesion boundary verification and classification resiliency. Table 5 shows an ablation study with the Mendeley dataset.

Statistical validation

To statistically validate the observed performance gains, we conducted a 5-fold paired t-test comparing the complete GRIFD model to its reduced variants. The t-test results showed that the whole model’s improvement in accuracy over the cross-pooling-removed version was statistically significant (p = 0.0032), as was its improvement in precision (p = 0.0047). Additionally, the 95% confidence intervals for GRIFD’s average accuracy and precision were [93.72%, 94.59%] and [95.75%, 96.35%], respectively, indicating high reliability and low variance in performance. These findings confirm that the enhancements introduced by GRIFD are not only consistent but also statistically meaningful.

Summary

The comparison of performances in Tables 6 and 7 emphasizes the better performance of the proposed GRIFD model compared to both existing methods (RTNet, MCNN-UNet, DRFEC) and conventional deep learning baselines (CNN, LSTM, CNN-LSTM, Transformer). On the feature level, GRIFD obtained the best accuracy (94.35%), precision (96.05%), and specificity (94.91%), while having the smallest mean error and detection delay. Equivalently, at the edge level, GRIFD also outperformed with 93.72% accuracy and 96.17% precision. Baseline models such as CNN and LSTM, on the other hand, demonstrated considerably poor specificity and detection times. These findings validate that GRIFD’s integration of cross-pooling, RNN sequencing, and ROI-based dissection greatly contributes to both consistency in segmentation and reliability in lesion classification across different areas in fundus images. Table 6 compares GRIFD with RTNet, MCNN-UNet, and DRFEC based on five evaluation metrics (accuracy, precision, specificity, mean error, and detection delay) using features as the basis for analysis.

The proposed GRIFD improves accuracy by 8.25%, precision by 8.12% and specificity by 12.24%. This method reduces mean error by 11.23% and detection delay by 9.04%. Table 7 provides a similar comparison as Table 6 but evaluates the metrics based on edge features, confirming GRIFD’s superiority in structural and boundary-based analysis.

The proposed GRIFD improves accuracy by 11.25%, precision by 9.3% and specificity by 10.98%. This method reduces mean error by 8.49% and detection delay by 9.69%. The observed improvements in precision and specificity can be directly attributed to GRIFD’s modular innovations. The graded ROI dissection improved lesion localization, while the cross-pooling module filtered spatial noise. Meanwhile, the RNN layer provided contextual continuity. These contributions collectively reduce misclassification and enhance decision reliability, as confirmed by the ablation results.

Limitation

Although GRIFD generally exhibits good performance, some instances of failure were observed. Misclassifications were primarily seen in low signal-to-noise ratio images, motion-blurred images, or in images where lesions become indistinguishable from the vascular tissue. Moreover, when trained on very imbalanced datasets like EyePACS, the model was found to have slightly lower sensitivity in minority classes. These mistakes imply that future research should incorporate sophisticated data balancing methods and adaptive attention components to enhance robustness. Another restriction exists in the sequential RNN inference, though optimised, which may be outperformed by parallelised attention-based options for increased speed and lesion boundary sensitivity.