ResNet, introduced by He et al. [9], has been widely adopted for AD detection because of its ability to train deep networks without suffering from the vanishing gradient problem. An improved ResNet model was proposed for the early diagnosis of AD via MRI scans [25]. This model, which is based on ResNet-50, incorporates several enhancements, including the Mish activation function, spatial transformer network (STN), and a nonlocal attention mechanism. These improvements enabled the model to capture long range correlations in MRI data while retaining critical spatial information, addressing the limitations often encountered by traditional CNNs. The model achieved a classification accuracy of 0.97 with the ADNI dataset, surpassing other algorithms in terms of macro precision, recall, and F1 score.

Additionally, a novel multistage deep learning framework based on residual functions was introduced for AD detection [26]. Inspired by the success of ResNet in image classification tasks, the framework employs five stages to enhance feature extraction while maintaining depth. Following the feature extraction phase, machine learning classifiers such as support vector machines (SVMs), random forests (RFs), and software were applied for classification. The model achieved excellent accuracy for three benchmark datasets (ADNI, MIRAID, and OASIS), with accuracy rates reaching as high as 0.99, outperforming existing systems.

Recent advancements have also highlighted the potential of RegNet in medical imaging applications, including AD detection. In one study [3], RegNet X064 was employed to predict amyloid deposition in PET scans for AD prognosis, achieving notably high performance. When combined with gradient boosting decision trees (GBDTs), the model exhibited reduced error margins and faster prediction times than human experts did. These findings underscore the effectiveness and scalability of RegNet in AD imaging tasks, making it a promising tool for clinical use in detecting neurodegenerative diseases.

DenseNet has also shown significant promise in the field of medical imaging, particularly in the analysis of MRI brain scans. Compared with traditional CNNs, DenseNet’s densely connected convolutional architecture facilitates more efficient feature extraction with fewer parameters, leading to improved performance in medical image analysis tasks [27]. In brain MRI analysis, DenseNet has demonstrated high accuracy in capturing intricate brain structures, making it a valuable tool in both clinical applications and research settings. This model’s ability to handle complex medical images efficiently demonstrates its potential to increase diagnostic accuracy and support advancements in neuroimaging techniques.

Several studies have explored the use of DenseNet in AD classification tasks. For example, in [28], a transfer learning based model utilizing DenseNet was introduced for classifying AD into three categories. This model achieved an accuracy of 0.96 and an AUC (Area Under Curve) of 0.99 with MRI datasets. This study demonstrated that DenseNet outperforms other traditional models in managing high dimensional MR data, particularly when combined with data augmentation techniques, addressing the issue of limited dataset availability. Moreover, the integration of a healthcare decision support system (HDSS) alongside the DenseNet model provided valuable insights for clinical decision making. These advancements highlight DenseNet’s potential to improve diagnostic accuracy in AD classification within clinical settings.

EfficientNet has also emerged as powerful deep learning model for medical imaging tasks, particularly those involving MRI scans. For example, in [29], a study was conducted utilizing a fine-tuned EfficientNet architecture for brain tumor classification that achieved superior performance across multiple datasets. EfficientNet’s efficient feature extraction and reduced computational complexity have suggested to be highly beneficial for analyzing high resolution MR images, particularly in complex brain imaging tasks. This study underscores the potential of transfer learning based EfficientNet models to increase diagnostic accuracy in medical imaging applications, as these models outperform state-of-the-art methods.

In AD related tasks, recent studies have demonstrated the effectiveness of EfficientNet. For example, EfficientNet-B0 was employed to classify brain MR images for early AD detection [30]. This approach integrates UNet for brain tissue segmentation and EfficientNet-B0 for feature extraction and classification. This model achieved an accuracy of 0.98, with high sensitivity and precision scores. These findings demonstrate EfficientNet’s ability to handle the complexity of brain MRI data, particularly in distinguishing between healthy and diseased brain tissues. The integration of EfficientNet into AD diagnosis systems has the potential to increase diagnostic accuracy and support early intervention, aligning with the growing body of research that positions EfficientNet as a robust tool for neurodegenerative disease classification.

While deep learning models such as CNNs have significantly advanced medical imaging tasks such as brain MRI classification, the interpretability of these models remains a critical challenge. However, techniques such as GradCAM have emerged as powerful tools to increase the transparency and interpretability of CNN models by generating visual explanations of the regions in an image that contribute most to a model’s output. In one study [31], the effectiveness of GradCAM in interpreting CNN models trained to classify different types of multiple sclerosis (MS) was demonstrated using brain MR images. The results showed that GradCAM provided superior localization of discriminative brain regions, making it an invaluable tool for understanding CNNs’ decision making processes in medical contexts. These results emphasize the importance of integrating interpretability techniques such as GradCAM to improve the reliability and clinical applicability of deep learning models in medical imaging.

In the context of AD diagnosis, recent studies have explored the combination of deep learning models with interpretability techniques such as GradCAM. For example, Inception ResNet a was applied to differentiate between AD patients and healthy controls (HCs) via T1-weighted MR brain images [32] and achieved competitive performance. GradCAM was employed to visualize the most discriminative brain regions, with the results indicating that the lateral ventricles in the mid-axial slice were key in distinguishing AD patients. This integration of GradCAM not only enhanced the transparency of the model’s decisions but also demonstrated its potential for assisting in diagnosis with minimal medical expertise. These findings highlight the importance of GradCAM in improving the interpretability and clinical relevance of deep learning models for AD diagnosis.

Krishnan et al. (2024) integrated a spatial attention mechanism into a CNN architecture to improve the classification of AD using MRI data. The spatial attention layer aids in guiding the model to focus on critical brain regions, such as those affected by AD, leading to a validation accuracy of 0.99. This approach, which assigns adaptive weights to important regions of the brain, highlights the potential of spatial focus techniques to increase both the accuracy and interpretability of deep learning models in AD diagnosis [33].

Sun, Wang, and He (2022) proposed a temporal and spatial analysis framework for improving AD diagnosis via the use of resting state functional MRI (fMRI) data. The authors employed a CNN with residual connections combined with a multilayer long short term memory network to classify AD patients and predict the progression of mild cognitive impairment (MCI) to AD. By constructing a functional connectivity (FC) network matrix based on regions of interest (ROIs) in the brain, the model was able to focus on critical brain areas while analyzing temporal changes over multiple time points. This approach aligns with the concept of spatial focus, as it directs the model’s attention to diagnostically relevant brain regions, enhancing the ability to distinguish AD patients from healthy controls and to predict progressive MCI (pMCI) patients vs. stable MCI (sMCI) patients. Their method achieved classification accuracy of 0.93 for AD patients vs. HCs and 0.75 for pMCI patients vs. sMCI patients, demonstrating the effectiveness of spatial and temporal focus in improving diagnostic accuracy [34].

The images are resized to 224 × 224 pixels to align with the standard input dimensions of various CNN architectures. This size ensures compatibility with pre-trained models, balances computational efficiency with feature preservation, and maintains anatomical integrity. Additionally, using a fixed resolution across all models eliminates input size variability, enabling a fair and consistent performance comparison.

To improve the generalizability of the models and prevent overfitting, several data augmentation techniques have been applied. These transformations introduce variations to the training dataset, allowing the models to learn more robust features. The following augmentations were used:

Random Horizontal Flip: With a probability of 0.50, each image is flipped horizontally. This transformation helps the model become invariant to the left–right orientation, which is especially useful in medical imaging tasks where structural symmetry is common.

These data augmentation techniques are crucial in enhancing the variability of the training data, which contributes to the robustness of the models during inference.

Once the augmentation techniques are applied, the images are directly converted to tensors via the transforms.ToTensor() function. This transformation function scales the pixel values (which originally range from [0, 255]) to a range of [0, 1] by dividing each pixel by 255, preparing the images for input into the neural networks.

After the images are converted to tensors, a normalization approach is applied using the mean and standard deviation values. Since the images are grayscale, the values used for normalization are specific to images with a single channel. The values used for normalization in this study are mean = 0.165 and standard deviation = 0.176.

Normalization ensures that the pixel values are scaled such that the resulting tensors have a mean of 0 and a standard deviation of 1. This process helps improve the convergence of the model during training by standardizing the input values, leading to more stable gradient updates and faster training.

The dataset is split into three subsets: training, validation, and testing. A stratified split is applied to ensure that the class distribution in the subsets reflects the overall class distribution in the dataset.

Training Set: 0.80 of the dataset is used for training the models.

This stratified split ensures a balanced distribution of classes across all subsets, maintaining a proportional representation of each class in the training, validation, and test sets.

The OASIS dataset presents a significant class imbalance, particularly between the nondemented and moderate dementia categories. To address this issue, we implemented a comprehensive approach that combines Weighted Random Sampling, a Weighted Loss Function, and Class-wise Performance Analysis to ensure that the model does not disproportionately favor the majority class.

To balance the dataset during training, Weighted Random Sampling was applied, ensuring that underrepresented classes were sampled more frequently. The class weights were computed using the formula: (1)ClassWeight=Nniwhere N represents the total dataset size, and ni is the number of samples in class i. This method ensured that all classes contributed equally during training, preventing the model from being biased toward the dominant nondemented category.

In addition to sampling adjustments, a Weighted CrossEntropy Loss Function was incorporated to further counteract the imbalance. The loss function was modified such that higher penalties were assigned to misclassified samples from underrepresented classes, ensuring that the model paid more attention to the minority classes. The loss function is formulated as follows: (2)L=−∑wcyclog⁡(yc)where wc represents the weight assigned to class c, which scales the contribution of each class in the loss calculation. This approach ensures that classes with fewer samples have a stronger influence on the optimization process, thereby improving the model’s ability to distinguish between dementia stages.

To evaluate the effectiveness of these techniques, a Class-wise Performance Analysis was conducted, where metrics such as precision, recall, F1 score, and accuracy were calculated separately for each class. This analysis validated that the model’s improvements were not driven solely by the majority class but were distributed more equitably across all dementia stages. The results confirmed that the integration of weighted sampling and loss functions led to more balanced predictions, reducing bias and improving overall classification stability.

A DataLoader is used to load the data in batches during the training process. The batch size is set to 32 across all the models, and the “WeightedRandomSampler” is incorporated into the training DataLoader to handle class imbalance. For the validation and test sets, the data are loaded without sampling, and shuffling is disabled to maintain consistency during evaluation.

(3)IPre=N(T(Rθ(Fp(Sd(Iorig)))),μ,σ)

IPre: Represents the image after applying all preprocessing steps.

In summary, the uniform preprocessing pipeline applied to all four models, RegNet, ResNet, DenseNet, and EfficientNet, serves as a cornerstone of this research. By ensuring that the same transformations and data preparation steps are consistently applied across all the architectures, we eliminate any potential biases that could arise from model specific preprocessing. This consistency guarantees that the results obtained from each model are directly comparable. Any observed differences in performance can thus be attributed solely to the inherent architecture of the models rather than to variations in the input preparation. This careful control of preprocessing conditions is a critical contribution of our work, ensuring a fair and rigorous evaluation of each model’s capabilities.

The unified preprocessing pipeline can be expressed mathematically as follows: (7)IPre=N(T(Rθ(Fp(Sd(Iorig)))),μ,σ)where:

IPre represents the image after all preprocessing steps.

Following this preprocessing pipeline, the image is passed through each of the four models: (8)IoutRegNet=RegNet(IPre) (9)IoutResNet=ResNet(IPre) (10)IoutDenseNet=DenseNet(IPre) (11)IoutEfficientNet=EfficientNet(IPre)

IoutModel represents the final output from each model.

By applying this uniform preprocessing pipeline followed by each specific model architecture, we ensure that any observed differences in performance are a direct result of the model’s architecture, as the preprocessing steps are consistent across all the models.

ResNet [16] is a deep CNN that employs residual connections to address the vanishing gradient problem, enabling the training of deeper architectures without performance loss. This design makes ResNet particularly effective for tasks such as medical image classification.

In this study, a pretrained ResNet50 was fine-tuned to classify the stages of AD. The final fully connected layer was modified to output predictions for four classes, allowing the model to distinguish between different stages of dementia.

RegNet [14] is a CNN architecture designed to optimize network design spaces for improved scalability and performance. Unlike traditional architectures, RegNet employs a regularized approach to adjust model size and complexity, resulting in a more balanced and efficient network. This scalability makes RegNet particularly effective for a wide range of tasks, including medical image classification, while maintaining computational efficiency.

In this study, a pretrained RegNet-Y-400MF model was fine-tuned for AD classification. The final fully connected layer was adapted to output predictions for four distinct classes, allowing the model to effectively differentiate between the various stages of dementia.

DenseNet [18] is a CNN that establishes dense connections between each layer to every other layer in a feedforward manner, ensuring maximum information and gradient flow throughout the network. This dense connectivity mitigates the vanishing gradient problem and improves feature reuse, leading to efficient learning with fewer parameters. Its ability to enhance feature propagation makes it particularly effective for tasks such as medical image classification.

In this study, a pretrained DenseNet121 model was fine-tuned to classify AD stages. The original classifier layer was replaced to output predictions for four classes, enabling the model to effectively distinguish between different stages of dementia.

EfficientNet [20] is a CNN that scales the dimensions of depth, width, and resolution in a balanced manner via a compound scaling method. This approach enables the network to achieve better accuracy and efficiency than traditional models do, making it highly effective for tasks requiring computational efficiency, such as medical image classification.

In this study, a pretrained EfficientNet-B0 model was fine-tuned to classify AD stages. The original classifier layer was modified to output predictions for four classes, allowing the model to accurately differentiate between the various stages of dementia.

For each model (ResNet, RegNet, DenseNet, and EfficientNet), the final fully connected layer was replaced to output four classes corresponding to AD stages (mild dementia, moderate dementia, nondemented, very mild dementia). This modification ensures that each model can classify input images into one of these four stages.

In all cases, the output layer was followed by a softmax activation function to convert logits into class probabilities, ensuring that the sum of the probabilities was 1 for each prediction.

The key training parameters applied consistently across all CNN models (ResNet, RegNet, DenseNet, and EfficientNet) are summarized in Table 2.

To ensure a fair comparison between the models, the following training parameters were applied consistently across all models, with the only difference being the loss function used in the EfficientNet-B0 model (Table 3).

The key distinction is that while the baseline models (ResNet, RegNet, DenseNet, and EfficientNet) utilize CrossEntropyLoss with class weights to manage class imbalance, the EfficientNet-B0 model employs CrossEntropyLoss with dynamic weighting based on GradCAM.

GradCAM was applied to the final convolutional block of EfficientNet-B0 before the classifier to generate heatmaps that highlight the most relevant image regions for each class prediction. These heatmaps guided the dynamic weighting in the loss function, ensuring the model focuses on the most critical regions of the image during training. The 3D GradCAM visualization further highlights the most significant brain regions involved in the classification process, providing a crucial understanding of the spatial focus in the model’s predictions Fig. 3.

For each input image, GradCAM generated heatmaps that highlighted the most relevant features for classification. These heatmaps were computed from the gradients of the output with respect to the activation maps in the target layer. The activations were weighted by the gradients and averaged spatially to produce the heatmap, which indicates the importance of each region in the input image for the model’s decision.

The GradCAM heatmaps were averaged across spatial dimensions to compute scalar weights representing the importance of the highlighted regions. These weights were then applied to the model outputs to scale them, ensuring that the network focused more on the regions of greater importance. This process allows the model to adapt its predictions by focusing on the most important parts of the image while ignoring less critical regions.

To enhance the model’s ability to focus on critical brain regions during training, GradCAM-based dynamic weighting was integrated into the CrossEntropy Loss function. The GradCAM heatmaps generated for each class were used to compute spatial importance weights, which were then applied to the model’s predictions before loss computation. This mechanism ensures that regions with higher clinical significance contribute more to the training process, allowing the model to refine its decision making based on essential anatomical features.

After each forward pass, the GradCAM activation maps were computed from the final convolutional layer. These heatmaps highlight the most influential areas in the brain MRI scans, with activation intensities reflecting their relative importance for each class.

For each class c, the weight wc, was computed by spatially averaging the heatmap activations, formulated as follows: (12)wc=1Z∑i,jαi,jcAi,jcwhere:

Ai,jc represents the activation at location (i,j) in the GradCAM for class c.

Before computing the loss, the predicted probability for each class y^c was adjusted dynamically using the computed GradCAM weight wc, ensuring the model prioritizes the most informative regions: (13)y^c weighted=wc⋅y^cwhere:

y^c is the raw model prediction for class c.

Using the modified outputs, the CrossEntropy Loss was calculated as: (14)Ldynamic=−∑cwc⋅yclog⁡(y^c weighted)where:

yc represents the ground-truth label for class c.

This integration ensures that the loss function penalizes errors more in clinically relevant regions, guiding the model to prioritize learning from the most diagnostically significant areas in MRI images.

The weights obtained from GradCAM represent the contribution of each activation map to the model’s final decision. These weights are used to determine the spatial importance of regions within the image.

The activation maps produced by the convolutional layers contain spatial information about the input image. Each pixel in the activation map corresponds to a specific region in the original image. By weighting the activation map using the gradients with respect to the class score, the regions that contribute more significantly to the model’s decision are highlighted. Regions with higher gradients are assigned greater importance in the final prediction, which is reflected in the GradCAM heatmap.

When GradCAM is used to guide the model during training, the following steps are applied:

GradCAM Calculation

After each training step, GradCAM is computed on the basis of the model’s current outputs. The gradients with respect to the target layer’s activations are calculated and used to generate the heatmap.

By incorporating GradCAM derived weights during training, the model is guided to focus on the most critical regions of the input image for classification, enhancing its ability to generalize and make accurate decisions. This approach is particularly beneficial in tasks where spatial information is crucial for understanding the image content, as the model can concentrate on the most relevant spatial features.

Table 4 highlights the performance of different models in classifying the mild dementia category. Among these models, EfficientNet with Dynamic GradNet yielded the highest overall F1 score (0.9881), reflecting a balanced trade off between precision and recall. However, EfficientNet achieved the highest accuracy (0.9976), indicating its superior ability to correctly classify both positive and negative samples in this category.

In the moderate dementia class, DenseNet achieved the highest accuracy (0.9999) and F1 score (0.9910), indicating its strong performance in correctly identifying both true positives and true negatives. Moreover, EfficientNet with Dynamic GradNet achieved the highest precision (0.9946), highlighting its robustness in minimizing false positives in this class, as shown in Table 5.

For the nondemented class, EfficientNet with Dynamic GradNet achieved the highest recall (0.9950), F1 score (0.9949), and accuracy (0.9953), making it the most effective model for correctly identifying nondemented individuals. ResNet and EfficientNet exhibited the highest precision values (0.9973 and 0.9971, respectively), indicating their strong ability to minimize false positives in this category, as shown in Table 6.

In the very mild dementia class, EfficientNet with Dynamic GradNet outperformed all the other models, with the highest precision (0.9779), F1 score (0.9773), and accuracy (0.9767). This finding indicates the superior ability of EfficientNet with Dynamic GradNet to accurately detect early-stage dementia while maintaining a lower rate of false positives, as shown in Table 7.

Table 8 presents a summary of the average performance of each model across all classes. EfficientNet with Dynamic GradNet yielded the highest average scores across all the metrics, including precision (0.9896), recall (0.9878), F1 score (0.9887), and accuracy (0.9903). This finding indicates that EfficientNet with Dynamic GradNet not only performs well across individual classes but also generalizes effectively across the entire dataset as illustrated in Fig. 5.

The integration of EfficientNet with Dynamic GradNet has notably enhanced the model’s ability to focus on critical regions, as illustrated in Fig. 6. The GradCAM visualizations demonstrate a progressive refinement in attention across four stages of Alzheimer’s disease progression (moderate dementia, mild dementia, very mild dementia, and nondemented). Each stage is represented by a sequence of four images, showing how the model’s focus improves over training, gradually capturing more relevant brain regions in MRI scans. Additionally, Fig. 6, also presents examples of failure cases from other models, which failed to emphasize critical areas and missed essential brain regions relevant to Alzheimer’s diagnosis. These failure cases (not EfficientNet with Dynamic GradNet) highlight the limitations of models without dynamic weighting, further emphasizing the effectiveness of the proposed approach in improving model robustness and reliability for MRI-based Alzheimer’s disease classification.

The results for the mild dementia class are further detailed with standard deviations, highlighting the consistency of each model. EfficientNet with Dynamic GradNet not only achieved the highest precision and F1 score but also maintained a low variance, indicating its stable and reliable performance across different test samples.

Table 9 reports the standard deviation (STD) for each metric, providing insight into the consistency and reliability of the model’s performance. The low variance across multiple runs further supports the robustness of the proposed EfficientNet with Dynamic GradNet framework.

Table 10 presents a comparison of each model’s sensitivity (recall) and specificity across the four classes. EfficientNet with Dynamic GradNet consistently performed well across all categories, particularly in the nondemented and moderate dementia classes with high specificity, indicating its strong ability to minimize false positives. Similarly, it achieved high sensitivity in detecting true positives, especially in the very mild dementia class, which is crucial for early diagnosis. The improvements in specificity and sensitivity achieved by EfficientNet with Dynamic GradNet compared to other models, including DenseNet, RegNet, and the baseline EfficientNet, are comprehensively illustrated in Figs. 7 through 14.

In the mild dementia class, EfficientNet with Dynamic GradNet achieved the highest F1 score (0.9881), indicating a well-balanced performance between precision and recall. This outcome suggests that the proposed model is highly effective in maintaining a low false positive rate while still capturing the majority of true positives. Interestingly, EfficientNet achieved the highest accuracy (0.9976), indicating its proficiency in correct classification overall. However, the relatively low F1 score (0.9799) than that of EfficientNet with Dynamic GradNet suggests that EfficientNet may exhibit a slight imbalance between precision and recall, possibly leading to a greater number of false negatives in this category.

The consistently high performance of EfficientNet with Dynamic GradNet can be attributed to the integration of the Dynamic GradNet mechanism, which likely enhances the model’s robustness by dynamically adjusting gradient updates, leading to more stable and accurate predictions. The marginal yet meaningful improvements in precision and recall over other models highlight the advantage of this approach in handling the complex and nuanced features associated with early-stage dementia.

For the moderate dementia class, EfficientNet with Dynamic GradNet achieved the highest F1 score (0.9949), demonstrating a well-balanced performance between precision and recall. While DenseNet attained the highest accuracy (0.9999), F1 score remains a more comprehensive indicator of the model’s robustness, particularly in cases with class imbalance.

The results suggest that while DenseNet excels in capturing both positive and negative samples with high accuracy, EfficientNet with Dynamic GradNet may be more suitable when the primary concern is reducing false positive rates, which is often a priority in clinical settings. This tradeoff between accuracy and precision must be carefully considered when selecting a model for deployment in real-world diagnostic systems.

In the nondemented class, EfficientNet with Dynamic GradNet demonstrated superior performance across multiple metrics, achieving the highest recall (0.9950), F1 score (0.9949), and accuracy (0.9953). This finding underscores its ability to accurately identify individuals who are not suffering from dementia, with minimal false negatives and false positives. High recall in this category is crucial, as misclassifying nondemented individuals could result in unnecessary anxiety and further diagnostic procedures.

Interestingly, ResNet and EfficientNet achieved the highest precision values (0.9973 and 0.9971, respectively), which indicates their effectiveness in minimizing false positives in this class. However, their lower recall values than that of EfficientNet with Dynamic GradNet suggests a potential shortcoming in correctly identifying all nondemented individuals, possibly leading to a greater number of false negatives. This trade-off between precision and recall is again evident, with EfficientNet with Dynamic GradNet offering a more balanced approach in this category.

Early detection of dementia is crucial for timely intervention, and in the very mild dementia class, EfficientNet with Dynamic GradNet outperformed all other models in terms of precision (0.9779), F1 score (0.9773), and accuracy (0.9767). This finding highlights the model’s ability to accurately detect early-stage dementia while maintaining a low false positive rate. Given the subtle nature of early dementia symptoms, this performance is particularly noteworthy, as it demonstrates the model’s ability to distinguish between very mild cognitive impairment and normal aging processes.

Moreover, EfficientNet with Dynamic GradNet exhibited exceptional performance in the very mild dementia category, surpassing all other models in terms of precision, recall, and F1 score. Its remarkable ability to detect early-stage dementia with minimal false positives and near perfect sensitivity makes it a highly reliable tool for early diagnosis. The model’s superior F1 score highlights its robustness in accurately identifying true cases while maintaining a low misclassification rate. As a result, EfficientNet with Dynamic GradNet stands out as the most powerful and efficient model for detecting this crucial early-stage of dementia.

The relatively low performance of other models in this class, particularly ResNet F1 score (0.8527) and RegNet F1 score (0.8893), suggests that these architectures may struggle with the fine-grained distinctions required for early dementia classification. The enhanced performance of EfficientNet with Dynamic GradNet can likely be attributed to its dynamic gradient adjustment mechanism, which allows for better generalizability in categories with subtle and overlapping features.

When comparing the models’ average performance across all classes, it is evident that EfficientNet with Dynamic GradNet stands out, achieving the highest average precision (0.9896), recall (0.9878), F1 score (0.9887), and accuracy (0.9903). This finding suggests that the proposed model not only excels in individual categories but also generalizes effectively across the entire dataset. The consistency in its performance, coupled with low variance in key metrics, indicates that EfficientNet with Dynamic GradNet offers a highly reliable solution for multiclass dementia classification.

The superior performance of EfficientNet with Dynamic GradNet across all classes can be attributed to its ability to adapt dynamically to different data distributions and class imbalances, which is a common challenge in medical datasets. The model’s ability to maintain high sensitivity (recall) and specificity across all categories further reinforces its potential as a robust tool for clinical applications, where both high true positive rates and low false positive rates are critical.

The sensitivity (recall) and specificity analysis across classes further highlights the strengths of EfficientNet with Dynamic GradNet. For the nondemented and moderate dementia classes, the model achieved the highest specificity values (0.9991 and 0.9964, respectively), indicating its strong ability to minimize false positives.

This balance between sensitivity and specificity underscores the clinical applicability of EfficientNet with Dynamic GradNet, as it has the ability to correctly identify true positives while minimizing false positives, thereby reducing the risk of misdiagnosis and unnecessary treatments.

The proposed EfficientNet with Dynamic GradNet model consistently outperformed existing architectures across multiple dementia classification tasks. Its superior precision, recall, F1 score, and accuracy—coupled with its ability to generalize well across different classes—make it an ideal candidate for clinical deployment in AD diagnosis. The model’s dynamic gradient adjustment mechanism allows it to handle the inherent complexities and imbalances in medical datasets, ensuring both high sensitivity and specificity. These findings suggest that EfficientNet with Dynamic GradNet offers a valuable contribution to the field of medical image analysis, particularly in the early detection of dementia, where accurate and timely diagnosis is paramount.