ARMDiaRD: A robust multi-class diabetic retinopathy detection using hybrid swin transformers with hierarchical fusion

Deep learning architectures that use fundus images in DR detection have shown good results in automating the identification of various stages of DR, but challenges persist. In particular, the efficient handling of large datasets and the accurate classification of severity grades remain areas that require further research. As models become deeper, the risk of missing important characteristic correlations increases, and the model becomes overfitted, potentially degrading classification performance when applied to a new set of data [12]. Thus, extracting highly sensitive deep features from extensive datasets for severity classification remains a critical challenge.

Motivated by the identified challenges, this research proposes ARMDiaRD: a novel hybrid Swin Transformer model combined with improved marine predator optimization (IMPO) for multiclassification of DR severity grades with a massive integrated global fundus dataset. The study also introduces global context (GC) capture with local feature concatenation using multi-scale feature fusion (MSFF), which enables deeper feature extraction and improves classification outcomes.

The primary contributions of this study are listed below.

• Introduction of the hybrid model, which first enriches the convolutional neural network (CNN) features with GC attention through EfficientNet-B0 and feeds them into Swin layers that deploy MSFF.

• Development of multiple Swin block architectures with MSFF for depth feature extraction, thereby enhancing classification performance by concatenating multiple features extracted from large datasets.

• Use of a hyperparameter tuning algorithm, improved marine predator algorithm (IMPA), for better performance through searching for the optimized hyper-parameter space.

• Extensive experimentation and result discussion carried out using the existing multiclassification fundus benchmark datasets, EyePACS, APTOS, MESSIDOR-V1, MESSIDOR-V2, and IDRiD. The proposed model has outperformed state-of-the-art models in comparison.

This paper is structured as follows. Section II provides an overview of the existing methods for the detection of DR and studies of various fundus data sets. Section III describes the proposed system methodology. Section IV presents the experimental setup and results, and Section V concludes the paper with future research hopes in the area of DR.

Fundus photography is the cornerstone of the diagnosis and classification of DR, enabling non-invasive visualization of the retina to identify pathological features such as microaneurysms, hemorrhages, and exudates. Numerous annotated fundus image datasets have been developed to facilitate automated DR detection using deep learning approaches, as given in Table 1 with their image dimensions and nativity. The EyePACS dataset, popularized through a 2015 Kaggle competition, continues to be one of the largest publicly available DR datasets, comprising over 88,000 images, with 35,126 labeled fundus images annotated with five classes according to the ICDR DR grades [19]. APTOS 2019 dataset includes 3,662 preprocessed images with the same grading scale and was designed for high consistency in classification tasks [20]. The IDRiD dataset [21] is unique in offering both DR severity labels and fine-grained pixel-level segmentation of lesions, including microaneurysms, hemorrhages, hard exudates, and soft exudates. FGADR dataset [22] provides fine-grained annotations and segmentation masks over 1,842 high-resolution images, supporting advanced lesion-aware DR grading research. MESSIDOR, released in 2008, includes 1,200 images graded for DR severity and image quality [23]. Although it uses a 4-class grading scheme, it is often mapped to the 5-class ICDR system for consistency. The diabetic retinopathy dataset (DDR) from China has 13,000 images with both DR and diabetic macular edema (DME) labels. The E-Ophtha dataset [24] provides images with pixel-level annotations for microaneurysms and exudates for lesion-focused research, and it lacks full DR severity grades. Together, these datasets form the foundation of modern DR classification and segmentation research, offering diverse combinations of image resolutions, annotation types, and patient demographics. They are frequently used to train and validate CNNs, Vision Transformers, and hybrid deep learning architectures.

The state-of-the-art literature is reviewed across benchmark datasets, such as APTOS, EyePACS, IDRiD, ODIR, and MESSIDOR, as shown in Table 2, done across conventional CNNs, hybrid models, vision transformers, and Swin transformer models. The CNNs have played a pivotal role in advancing computer vision by solving various visual recognition challenges. A CNN is structured with multiple convolutional layers for feature extraction, each designed to learn and recognize specific features or patterns in the input image. These layers compute feature maps by sliding small filters across the image. Pre-trained CNN models may be further fine-tuned for different applications, leveraging knowledge gained from exposure to large, diverse datasets. Transfer learning is a process of fine-tuning pre-trained CNNs, such as VGGNet, ResNet, Inception, and DenseNet, on the targeted fundus dataset for further learning and feature selection. Early approaches relied heavily on CNN architectures, such as Inception-v3 and ResNet [25]. Gulshan et al. [19] pioneered a high-accuracy binary classification system for DR detection using Inception-v4 on Eye-PACS, reporting an AUC of 0.99. Porwal et al. [21] established the first benchmark using Inception-v3 on the IDRiD dataset, achieving around 78% accuracy. Islam et al. (2020) [26] advanced this by introducing a multitask CNN that jointly performed lesion segmentation and classification, attaining a Quadratic Weighted Kappa (QWK) score of 0.80. Pratt et al. (2016) [27] employed a 10-layer CNN to perform 5-class classification on EyePACS with approximately 75% accuracy. Other enhancements, such as the use of DenseNet-121 with focal loss (Wang et al., 2021) [28] and ResNet50 with domain adaptation (Voets et al., 2019) [29], tackled data imbalance and cross-domain generalization issues. Another research conducted by Tiwari et al. [30] also claimed that the MobileNet outperformed all the other architectures while testing in the fundus photography dataset [30] when tested with EyePACS. Dai et al. [38, 19] introduced DeepDR Plus, a deep learning system developed to estimate the time to progression of DR, a major cause of preventable vision loss. The model was pretrained on an extensive dataset comprising 717,308 fundus images from 179,327 individuals and validated using 118,868 images from a multiethnic cohort of 29,868 participants. DeepDR Plus utilized a pretrained ResNet-50 as a feature extractor, complemented by using self-attention to focus on diagnostically relevant areas in fundus images. It achieved strong predictive performance, with concordance indexes between 0.754 and 0.846. Although the model was trained primarily on a Chinese population, its generalizability could be enhanced through further training using more diverse clinical and demographic datasets.

Hyperparameter tuning using optimization algorithms helps in achieving good performance for deep learning models. Manual tuning or grid search is often inefficient, especially when the search space is large or when training is computationally expensive. To address this, optimization algorithms are employed to automate the hyperparameter search process more effectively. The improved marine predators algorithm (IMPA) is a population-centric optimization method, shares similarities with various other metaheuristic strategies, particularly with the enhanced marine predators algorithm (MPA) [41, 42]. A notable attribute of this algorithm is its periodic dispersion of the starting solution throughout the search domain, as articulated in Eq. (1). (1) X0=Xmin+rand⋅Xmax−Xmin {X_0} = {X_{\min }} + {\rm{rand}} \cdot \left( {{X_{\max }} - {X_{\min }}} \right) where rand denotes a randomly generated vector with elements uniformly distributed between 0 and 1, while X_min and X_max represent the minimum and maximum bounds of the parameters being optimized, respectively. This initialization ensures that the initial population of candidate solutions is spread uniformly across the feasible search space, laying a solid foundation for the subsequent optimization phases. Each candidate in the population is represented as a vector of hyperparameter values, for example, Xi=learning rate,batch size,depth,dropout,…. {X_i} = \left[ {{\rm{learning\; rate}},{\rm{ batch\; size}},{\rm{ depth}},{\rm{ dropout}},\; \ldots } \right].

For each candidate X_i, the model is trained using the corresponding hyperparameters (HP). After training, the model is evaluated on a validation set to obtain a performance metric. For classification tasks, commonly used metrics include validation accuracy or F1-score. This validation performance metric is then used as the fitness value for the candidate, depending on the problem X_i as mentioned in Eq. (2). (2) fitness=validation accuracy,if maximization problem−validation loss,if minimization problemand fitness maximizedvalidation loss,if minimization problemand fitness minimized {\rm{fitness}} = \left\{ {\matrix{{{\rm{validation}}\;{\rm{accuracy}},} \hfill & {{\rm{if}}\;{\rm{maximization}}\;{\rm{problem}}} \hfill \cr { - {\rm{validation}}\;{\rm{loss}},} \hfill & {\matrix{{{\rm{if}}\;{\rm{minimization}}\;{\rm{problem}}} \hfill \cr {{\rm{and}}\;{\rm{fitness}}\;{\rm{maximized}}} \hfill \cr } } \hfill \cr {{\rm{validation}}\;{\rm{loss}},} \hfill & {\matrix{{{\rm{if}}\;{\rm{minimization}}\;{\rm{problem}}} \hfill \cr {{\rm{and}}\;{\rm{fitness}}\;{\rm{minimized}}} \hfill \cr } } \hfill \cr } } \right.

IMPA finds the optimized hyperparameter solutions faster by using certain social dynamics and dynamic weights initialization, which can prevent ending up with local optima. IMPA is better at exploring global optima, improving overall performance. The update population X in IMPA update rules uses Levy flights and adaptive position adjustments inspired by marine predator behavior, balancing exploration and exploitation in the search space.

ARMDiaDR system combines multiple data sources captured using different devices or imaging conditions. To gain a more comprehensive view, a strong data agglomeration process is carried out on the individual datasets before the model training process to enhance both image quality and model training efficiency. The entire pipeline consists of three main phases: preprocessing, augmentation, and quality-aware down-sampling. This process serves as an important feature, as images from various datasets (APTOS, IDRID, MESSIDOR-V1, MESSIDOR-V2, and EYEPACS) are merged for experimentation. The pipeline automatically handles the preprocessing of multiple datasets as shown in Figure 2, eliminating the need for manual intervention and ensuring consistency across datasets. The images are loaded from their respective datasets, which are preprocessed and aggregated, and the majority classes are subject to quality-aware down-sampling and augmented to generate the training and testing datasets.

Figure 2:

Data agglomeration pipeline with sample images showing the three stages of data preprocessing, quality-aware down-sampling, and augmentation ending with splitting for the training and testing process.

The quality-based down-sampling strategy was incorporated to ensure that only diagnostically rich images were retained, particularly among healthy samples, where overrepresentation is common for generating the balanced dataset. Image quality was assessed using a set of objective metrics, including sharpness (edge clarity), contrast (luminance and color differentiation), brightness (global illumination), color saturation (intensity of color representation), and noise level (random pixel intensity variations).

These metrics were computed for each healthy image, allowing a quality-based ranking of the dataset. Down-sampling was then performed by selecting high-scoring images, thereby ensuring that the reduced dataset preserved clinically meaningful features.

a.ii

Quality-aware down-sampling

To ensure a balanced dataset suitable for training machine learning models, a quality-aware down-sampling strategy is employed on each dataset, as the fundus datasets are highly imbalanced. A threshold value of 7,000 images is set for the combined dataset for each class, consolidating to 35,000 images for all five classes. Each image is scored using a composite quality metric defined in Eq. (3), which combines sharpness, contrast, and brightness. (3) Q=0.5⋅Sharpness+0.3⋅Contrast+0.2⋅Brightness Q = 0.5 \cdot {\rm{Sharpness}} + 0.3 \cdot {\rm{Contrast}} + 0.2 \cdot {\rm{Brightness}} where

• Sharpness = variance of Laplacian [45],

• Contrast = standard deviation of pixel intensities.

• Brightness = mean pixel intensity.

Images with a low Q score are excluded from the training dataset. This filtering process improves the signal-to-noise ratio and ensures that the model focuses on diagnostically useful information. Quality-aware filtering is particularly beneficial when training architectures, such as Swin Transformers, which are sensitive to feature representation quality. The histogram results shown in Figure 3 show how the quality of the original images varies with the down-sampled balanced images with respect to the Q-scores of the images selected for training in each class of DR grading. Table 3 shows the number of balanced images after the data agglomeration pipeline used for experimentation of the proposed system with its properties from the four datasets APTOS, IDRID, MESSIDOR V1&V2, and EYEPACS.

Figure 3:

Histogram of Q-scores calculated for each grade of DR before and after quality-aware down-sampling, which reflects the improvement in quality of images selected for the oversampled classes. DR, diabetic retinopathy; ICDR, international clinical diabetic retinopathy.

Table 3:

textbfDataset	No DR	Mild NPDR	Moderate NPDR	Severe NPDR	PDR	Total Images	Dimensions
APTOS [20]	1,805	370	999	193	295	3,662	Mixed
IDRID [21]	168	25	168	93	62	516	4,288 × 2,848
EYEPACS [19]	25,802	2,438	5,288	872	708	35,108	Mixed
MESSIDOR-V1 [23]	546	153	247	254	–	1,200	2,240 × 1,488
MESSIDOR-V2 [46]	1,017	270	347	75	35	1,744	2,240 × 1,488
Training Dataset	5,600	5,600	5,600	5,600	5,600	28,000	224 × 224
Testing Dataset	1,400	1,400	1,400	1,400	1,400	7,000	224 × 224

DR, diabetic retinopathy; NPDR, non-proliferative diabetic retinopathy; PDR, proliferative diabetic retinopathy.

In the proposed system ARMDiaRD, the IMPA-driven tuning process [41, 42] is integrated to evaluate each hyperparameter configuration reliably. The use of such optimization algorithms significantly accelerates the convergence toward an optimal set of HP and enhances the overall predictive performance of the model. The improved marine predators algorithm (IMPO) is used for hyperparameter optimization of the proposed hybrid Swin Transformer model with MSFF. The fitness function within IMPO evaluates the effectiveness of different hyperparameter combinations, including the learning rate (lr), the embedding dimension (embed dim), and the dropout rate (dr). The pseudocode of the IMPA optimizer process is shown in algorithm 1.

These bounds ensure that the optimizer explores a diverse and feasible region of the hyperparameter space. Along with this, the population size was set to 10, and the upper limit on the number of iterations was set to 5. The hyperparameter bounds used in the experimentation is as follows:

• Learning Rate (lr): The search interval was set to [1 × 10⁻⁵, 1 × 10⁻³]. This controls the step size in gradient-based updates.

• Dropout Rate (dr): The search interval was [0.1, 0.5], which controls the fraction of neurons randomly dropped during training to prevent overfitting.

• Embedding Dimension (embed dim): This was searched within the range [48, 192] and defines the dimensionality of intermediate feature representations in the model.

A stratified five-fold cross-validation split was used to maintain class distribution consistency. For each candidate solution X_i, the model is trained using the generated HP within these bounds. The trained model is then evaluated on a validation set to obtain a performance metric, specifically the validation accuracy for classification tasks. The fitness value is calculated as shown in the algorithm 1 with the minimizing problem. The best-performing hyperparameter configuration for the learning rate, dropout rate, and embedding dimension across all iterations and candidates is recorded and used for the proposed ARMDiaRD system’s final training and testing process, as summarized in the results table.

The proposed ARMDiaRD system uses EfficientNet-B0, a baseline model, as its back-bone, initialized with ImageNet-1k pretrained weights. Built on it, a hybrid wiggled layers of Swin blocks with multiscale features fusion, as shown in Figure 4. The proposed system is a hybrid model that uses EfficientNet-B0, a lightweight and efficient CNN backbone, to extract initial image features as a baseline model. These features are enhanced using a GC block to capture broader spatial dependencies, followed by a 1 × 1 convolution layer for dimensional alignment. The core of the model consists of wiggled Swin Transformer blocks, each followed by a MSFF layer and global average pooling (GAP), enabling the model to capture both local and hierarchical global features. Outputs from each Swin block (GAP1, GAP2, and GAP3) are concatenated to form a rich feature representation, which is then passed to a fully connected layer for final classification.

Figure 4:

Proposed robust and hybrid deep learning model architecture with initially capturing the features using EfficientNet and GC block, followed by hybrid Swin Transformer blocks with MSFF and hierarchical aggregation. GC, global context; MSFF, multi-scale feature fusion.

For the experimental setup, a high-performance NVIDIA manufactured A100 80GB PCIe GPU, installed with PyTorch version 2.7.0 and CUDA version 12.2 from it's official website, is employed to ensure compatibility with the latest deep learning frameworks. During experimentation, the GPU operated at a stable temperature of 37°C with a power consumption of 67 W, well within its maximum capacity of 300 W.

To evaluate the proposed system ARMDiaRD, the evaluation metrics as listed in Table 4, with their importance and formulas, are adopted. Along with the normal metrics, such as accuracy, recall, precision, F1-score, the class-wise performance metric, QWK, and Spearman correlation, which considers the deviation from the original class in ordinal prediction problems, such as DR detection from no DR (Class 0) to PDR (Class 4), is also calculated. The class differences are measured by how far the false predictions fall with this metric, and QWK and Spearman correlation play a vital role in the proposed model evaluation.

Table 4:

Performance Metric	Mathematical Expression	Description
Accuracy	Accuracy=TP+TNTotal Instances {\rm{Accuracy}} = {{{\rm{TP}} + {\rm{TN}}} \over {{\rm{Total Instances}}}}	Provides a high-level overview of model performance.
Recall	Recall=TPTP+FN {\rm{Recall}} = {{{\rm{TP}}} \over {{\rm{TP}} + {\rm{FN}}}}	Measures how many actual positive instances the model correctly identified.
Precision	Precision=TPTP+FP {\rm{Precision}} = {{{\rm{TP}}} \over {{\rm{TP}} + {\rm{FP}}}}	Measures the accuracy of positive predictions.
F1-Score	F1-Score=2⋅(Precision⋅Recall)Precision+Recall {\rm{F}}1{\rm{ - Score}} = {{2 \cdot ({\rm{Precision}} \cdot {\rm{Recall}})} \over {{\rm{Precision}} + {\rm{Recall}}}}	Harmonic mean of precision and recall; balances false positives and false negatives.
QWK	K=Observed−Expected1−Expected K = {{{\rm{Observed}} - {\rm{Expected}}} \over {1 - {\rm{Expected}}}}	Measures agreement between two raters, penalizing bigger disagreements quadratically.
MAE	MAE=1n∑i=1nYi−Y^i {\rm{MAE}} = {1 \over n}\sum\limits_{i = 1}^n {\left| {{Y_i} - {{\hat Y}_i}} \right|}	Measures average magnitude of the errors between predicted and actual values.
MSE	MSE=1n∑i=1nYi−Y^i2 {\rm{MSE}} = {1 \over n}\sum\limits_{i = 1}^n {{{\left( {{Y_i} - {{\hat Y}_i}} \right)}^2}}	Similar to MAE but penalizes larger errors more by squaring them.
Spearman Correlation	ρ=1−6∑di2nn2−1 \rho = 1 - {{6\sum {d_i^2} } \over {n\left( {{n^2} - 1} \right)}}	Measures strength and direction of a monotonic relationship in ranked or ordinal data.

MAE, mean absolute error; MSE, mean squared error; QWK, Quadratic Weighted Kappa.

The hyperparameter tuning is performed for the experiment by a metaheuristic optimizer, the improved marine predator optimization (IMPO) algorithm, by training on the entire dataset for 10 epochs. Initially, the entire dataset is subjected to three nature-inspired optimizers, the improved marine predator optimization (IMPO) algorithm, the particle swarm optimizer (PSO), and the differential evolution (DE) to find the best optimizer for the given problem. The pair plot shown in Figure 5 visualizes the distribution of the three HP, learning rate lr, dropout dr, and embedding dimension embed dim explored during the exploration and exploitation process of the optimizers IMPO, PSO, and DE.

Figure 5:

Scatter plot of the HP lr, dr, and embed dim for the three optimizer algorithms, IMPO, PSO and DE. DE, differential evolution; dr, dropout rate; HP, hyperparameters; lr, learning rate; PSO, particle swarm optimizer.

The one-way analysis of variance (ANOVA) test is performed to determine the best optimizer based on accuracy, resulting in an F-statistic of 0.102 and a p-value of 0.904, indicating that there is no statistically significant difference among the three. However, IMPO is selected for intense hyperparameter tuning for the proposed system due to its nature of exploring a diverse range of options and achieving high accuracy results, as shown in Figure 6. Thus, the best optimized HP given by the IMPO optimizer fitness score listed in Table 5 are used for the training process.

Figure 6:

Accuracy box plot obtained from the three optimizer algorithms, DE, IMPO, and PSO. DE, differential evolution; PSO, particle swarm optimizer.

Five publicly available datasets, EyePACS, APTOS, MESSIDOR-V1, MESSIDOR-V2, and IDRiD, is used for training and validation analyses. The preprocessed, quality-aware, and augmented datasets comprise a total of 35,000 images is used in training, validation, and testing. The raw images were scaled to a resolution of 224 × 224 pixels and then normalized, and the agglomeration process was carried out as mentioned in Section 3.1 to improve the visibility of the lesion. The augmented dataset produced from the five various sources after the intense data agglomeration process was divided using an 80%–20% division for training and evaluation. To ensure reproducibility, we fixed the random seed across NumPy, PyTorch, and Python built-in libraries. By stratifying the folds, the five-fold cross-validation helps preserve the class balance in every fold, leading to a more reliable estimation of the generalization performance of the model. The generalized dataset used was trained and tested across various existing baseline classifier models listed in the cases below to validate the proposed system.

• Case 1: ResNet, the ResNet architecture of the baseline model (ResNet50) without any attention or transformer modules.

• Case 2: EfficientNet-B0 (EffiNet), the lightweight and efficient EfficientNet-B0 backbone, pretrained on ImageNet without any attention or transformer modules.

• Case 3: EffiNet with Global Context (EffiNet + GC), for contextual understanding, the GC block is added to the EfficientNet-B0 backbone.

• Case 4: EffiNet + GC and Squeeze-and-Excitation (EffiNet + GC + SE), building on the previous configuration (EffiNet + GC), introduced Squeeze-and-Excitation (SE) blocks for channel-wise attention. SE blocks adaptively recalibrate feature maps to emphasize informative features, thereby improving representational power.

• Case 5: EffiNet + GC and CBAM (EffiNet + GC + CBAM), this variant replaces SE, incorporating the Convolutional Block Attention Module (CBAM), which integrates both channel-wise and spatial attention.

• Case 6: EffiNet + GC with 3 Swin Transformer Blocks and Multiscale Feature Fusion (Proposed), the final model integrates three sequential Swin Transformer blocks following the EfficientNet backbone. Their outputs are fused using a MSFF mechanism, enabling the model to combine hierarchical transformer features adaptively across scales.

The comparison line chart shown in Figure 7 clearly shows that the proposed system has better performance compared to the various other models on the same dataset processed by the proposed system. The accuracy, precision, recall, and F1 score obtained during the training process in the five folds for the 40 epochs for the different cases are shown in Figures 8–12, with the proposed system in Figure 13. The proposed system has performed best in its training process with a weighted average accuracy of 87.42%, recall of 87.49%, and precision of 87.51% among all the cases mentioned.

Figure 7:

Results from stratified five-fold validation performed with average metrics of accuracy, precision, recall, and F1-score. CBAM, convolutional block attention module; GC, global context.

Figure 8:

Results from stratified five-fold validation performed during the training phase using Resnet-50.

Figure 9:

Results from stratified five-fold validation performed during the training phase using EfficientNet-B0.

Figure 10:

Results from stratified five-fold validation performed during the training phase using EfficientNet-B0 with GC. GC, global context.

Figure 11:

Results from stratified five-fold validation performed during the training phase using EfficientNet-B0 + GC + SE. GC, global context.

Figure 12:

Results from stratified five-fold validation performed during the training phase using EfficientNet-B0 + GC + CBAM. CBAM, convolutional block attention module; GC, global context.

Figure 13:

Results from stratified five-fold validation performed during the training phase using the proposed model ARMDiaRD.

The testing model for the proposed system ARMDiaRD is developed using an ensemble evaluation of all five models generated through the five-fold cross-validation training phase. The model achieved an accuracy of approximately 87.59%, with a precision of about 87.6% and a recall of around 87.9%. The QWK and Spearman coefficients, which help in analyzing the variation from the actual score in case of wrong prediction, are high, with 0.9147 and 0.9253. This shows the proposed hybrid Swin layers capture context at multiple scales and complement the GC with a channel attention pipeline. The important metric in medical diagnosis of false predictions calculated by the mean absolute error (MAE) drops from 0.3310 to 0.1736, reflecting a reduction in prediction error magnitude in the proposed system by adding a hybrid Swin Transformer with hierarchical fusion of features. Importantly, the MAE, a critical metric for evaluating false predictions in medical diagnosis, dropped significantly from 0.3310 (baseline) to 0.1736 in the proposed model. However, the integration of a hybrid Swin Transformer on top of the baseline model, combined with training on a large dataset, has ultimately outperformed all other configurations. The graph shown in Figures 14 and 15 gives the ablation study results with clear advantages in both classification accuracy and ordinal consistency (QWK, Spearman), while also minimizing the MAE. Overall, the proposed hybrid ARMDiaRD system integrating GC and channel attention into a Swin Transformer achieves the highest performance across all key metrics: accuracy, precision, recall, QWK, and Spearman correlation, along with the lowest MAE. These findings confirm that combining local, global, and hierarchical features within a carefully designed hybrid architecture is highly effective for robust medical image analysis.

Figure 14:

Results of ablation analysis performed with the various metrics, such as accuracy, precision, recall, QWK, Spearman correlation, and MAE with the training dataset subjected to the proposed agglomeration process, showing the overall improvement of the proposed model. MAE, mean absolute error; QWK, quadratic weighted kappa.

Figure 15:

Results obtained during the testing process with metrics, such as accuracy, precision, recall, QWK, Spearman correlation, and MAE. MAE, mean absolute error; QWK, quadratic weighted kappa.