3.1.1. ResNet

In this context, y signifies the output of the residual block, whereas H(.) function encompasses the stacking of layers within the residual function. The input to the block is denoted by x, while θ_k represents the network weights for given layers within the residual block, and θ_s corresponds to the skip connection network weights. To determine the successive input value x are obtained by applying the ReLU activation function to the sum of the weighted outputs from the previous layers. This application of ReLU usually introduces nonlinearity into the computation process.

3.1.2. DenseNet

The inherent characteristic of the DenseNet [49] architectures relies on feature reuse and the feedforward propagation prowess that allows the learning of robust features by leveraging information from previous layers, thus resulting in a compact network with fewer network parameters. The dense connections improve the backpropagation of gradients, resulting in improved speed of convergence and training stability. The creators of the architecture proposed four distinct versions, each sharing the same fundamental principles and consisting of 4 dense blocks having variations in the connectivity levels. The intrinsic portions of the initial two blocks were kept constant across all the model variants, with the first block comprising six dense layers and the second incorporating 12 dense layers. The dense layers in all versions employ two convolutions with kernel filter of sizes 3 × 3 and 1 × 1. Conversely, the arrangement of the final two dense blocks varies across different variants of this deep learning architecture. Our research investigation concentrated mainly on the four variations of the DenseNet architectures: DenseNet-201, DenseNet-169, DenseNet-161, and DenseNet-121.

3.1.3. EfficientNet

The research study by Tan and Le [50] introduced an innovative technique for concurrently optimizing the adjustment of the network width, image resolution, and network depth. The pioneer of this technique utilized the compound scaling paradigm while introducing a distinctive set of DL architectures, denoted as EfficientNet-B0 through B7. EfficientNet is constructed from foundational models that utilize Mobile Inverted Bottleneck (MBConv) blocks. Our present investigation examined three iterations of EfficientNet (EfficientNet-B7, EfficientNet-B4, and EfficientNet-B0).

3.1.4. MobileNetV3

3.2.1. Adam

At time step t, $$ approximates the gradients relative to the given loss function. 〈·(W_t−1) denotes the cross-entropy loss function, contingent upon the weight parameters W_t. The numerator segment of the fraction in this equation estimates the bias-corrected first moment β₁, while the denominator segment approximates the bias-corrected second raw moments β₂. Notably, the exponential decay rates for these moment estimates are raised to the power variable t, resulting in $$ and $$, respectively. Within an Adam optimization algorithm, updates are made to the squared gradient (V_t) and the moving mean of the gradient (m_t). In our preliminary experiments, we optimized the learnable weights on all the architectures (variants of ResNet, DenseNet, MobileNet, and EfficientNet) previously described in Section 3.1.

3.2.2. RAdam

In equation (3), the term inside the square root represents the variance rectification component. It is important to note that when the variance becomes tractable, ρ_t exceeds 4. Additionally, η_t represents an adaptive learning rate, which can be determined using the expression $$. Here, V_t+1 denotes an update for the estimation of the first bias moment and α accounts for the weight decay.

3.2.3. AdamW

AdamW is a stochastic optimization technique that adjusts the conventional approach to weight decay in Adam [54] by separating weight decay from the gradient update process. In Adam l₂ regularization is commonly integrated using the following modification, where W_t represents the weight decay rate at time t “with the following” The weight update in AdamW is formulated as: $$ where W_t denotes the weights at step t, η is the learning rate, m_t and v_t are the bias-corrected first and second moment estimates, ϵ is a small constant for numerical stability, and λ is the weight decay coefficient. This decoupling strategy improves regularization and enhances generalization performance across various deep learning tasks.

3.4.1. CAM

CAMs are useful for visualizing the decision rationale of DL models by producing heatmaps showing the model’s attention region for a given input image. This provides an intuition into the model’s decisions and enables researchers and end users to understand which regions of the input contribute to the classification result. CAMs are used for tasks such as object detection, image localization, and model debugging, helping to ensure the transparency and reliability of CNN-based models. Three CAM methods are considered in our study: GradCAM, GradCAM++, and AblationCAM.

3.4.2. SHAP

SHAP is an interpretability algorithm based on cooperative game theory utilizing the concept of Shapley values assigning a value for each player’s contributions in a game. The SHAP algorithm evaluates the feature contributions to the model’s output by considering all possible feature combinations and averaging the contributions to the output. In the field of XAI, the SHAP algorithm is popular due to the ability to provide global insights to model behavior. The SHAP algorithm is model agnostic and belongs to a class of additive feature attribution methods having these properties [29]: local accuracy, missingness, and consistency. The model-agnostic attribute makes it suitable for any machine learning or DL model. The SHAP algorithm is employed in our study for understanding the feature attributions of the developed models. We utilize the partition explainer function in the official SHAP API [29]. In all the explanations generated by the partition explainer, we utilized a total of 500 evaluations.

3.6.1. Design Stage

This stage entails conceptualizing and designing the frontend and backend of the AI-based logic for our proposed interpretable DL platform capable of predicting skin lesions.

3.6.2. Implementation Stage

In this stage, we implemented the user interface for our proposed interpretable DL system web application. The demo web application was built on a Python web-based library, Gradio, which supports modularized coding philosophy and facilitates the speedy deployment of AI models. The DL models and explainable algorithms were developed in Python to create a software platform facilitating man–machine interaction within the AI-enabled dermatology decision support platform. Additionally, Cascading Style Sheets (CSS) were employed to design the user interface of our application. Figure 4 shows the implementation steps for the proposed system.

3.6.3. Deployment Stage

We integrated three chosen DL models (DenseNet, ResNet, and MobileNetV3) into our application program backend in this stage. The visual interface of the AI system we created is depicted in Figure 5. We deployed the developed application on a local host.

4.1.1. Model Training With Random Weights

Table 4 summarizes the binary classification performance of the 13 different state-of-the-art architectures initially investigated in this study. All versions of DenseNet trained on original images surpass all other DL methodologies when assessed on the testing set (when attempting to solve the binary classification). They exhibit a performance ranging from 74.4% to 78.5% across various classification metrics such as accuracy, precision, recall, and F1-Score. The second best method concerning DL architecture is ResNet101, demonstrating notably superior performance to all instances of MobileNetV3 and EfficientNet variants. Conversely, MobileNetV3 variants exhibit the lowest performance levels. Additionally, we conducted a detailed analysis of the performance relative to each DL variant to identify the top-performing model within each architecture. Consequently, DenseNet161, ResNet101, EfficientNetB7, and MobileNetV3-L emerged as the top performers across their respective DL architectures. Therefore, we utilized the chosen DL techniques incorporating variations of the Adam optimizer (such as RAdam, Adam, and AdamW) to carry out supplementary experiments on the training set containing either original or data augmented images. We conducted additional experiments to assess whether there is potential for enhancing DL model performance factoring DA compared to DL methods, which are trained solely on original images.

The research outcomes were consolidated to gain insights into the binary classification problem, considering the optimization schemes previously described, and we subsequently trained the DL methods on the original and DA images. The results of the performance metrics (accuracy, F1-Score, and AUC) obtained after training the DL techniques and then evaluating the trained models on the testing set are summarized in Tables 5, 6, and 7. From these tables, we observed that DenseNet161 factoring at least one of the optimization schemes, considering the appropriate choice of the DA scheme, obtained the highest performance, outperforming other DL methods across the evaluated performance metrics. The DenseNet-161 architecture utilizing DA-1 and the Adam optimizer yielded the highest classification accuracy and F1-Score (79.4%), which outperformed the DenseNet161 model derived after training on original images (baseline) by a margin of 1%. To further understand the impact of the examined DA schemes, we inspect the 12 experimental scenarios: each scenario accounts (per row) for all possible experiments of a particular DL instance, factoring in one of the optimization schemes, and was trained on either original images or each of the DA schemes. Consequently, the models obtained after training were used to evaluate the testing set.

We report that the DA-1 scheme aided the DenseNet161 in obtaining the highest binary classification accuracy and F1-Score. Additionally, we inspected the impact of the DA across the 12 scenarios; the results show that DA-4 and DA-3 are the most influential in enhancing the DL model performance across the four main neural network architectures under study. The second competitive method is MobileNetV3-L architecture; the most favorable performance was observed with the DA-4 approach across all three optimizer types compared to all other DA schemes, while the Adam optimizer surpassed the AdamW and RAdam. Additionally, MobileNetV3-L, when trained on DA-4, significantly outperforms MobileNetV3-L when trained on the original image by a margin ≥10.3% across accuracy, F1-Score, and AUC metrics. Likewise, the best model accuracy was recorded for the EfficientNet-B7 architecture with DA-4 compared with all other model optimizers: RAdam performed slightly better than the other optimization schemes. The best models from the EfficientNet-B7 architecture outperform most variants of the ResNet101 in most of the performance evaluation metrics. After conducting experiments for the different instances of the ResNet-101 models, while considering different DA schemes and optimizers, the highest model accuracy of 76.2% was achieved with DA-3 when the ResNet101 model weights are optimized using the AdamW optimizer. In general, the ResNet-101 factoring DA-3 and DA-4 schemes yielded the best performance across the three main experimental scenarios for this architecture. However, the best-performing ResNet-101 model trained on data-augmented images still yielded poorer performance compared to most of the top-performing architectures, including DenseNet161, MobileNetV3-L, and EfficientNet-B7.

The results presented show that spatial or geometrical transformation–based DA schemes based on random rotation (DA-4), horizontal flipping (DA-3), and cropping (DA-1) can improve the binary classification performance of DL architectures that are trained on clinical skin lesion images. In most instances, the performance of these DL architectures deteriorates when color jitter DA (DA2) schemes are considered. It is vital to highlight that the random rotation (DA-4) yielded the best performances in most experiments, especially with the MobileNetV3-L and EfficientNet-B7. We underscore that some instances of DA-5, which factors DA-1, DA-3, and DA-4, often yielded decent high classification scores across the evaluated metrics. At the same time, the top-performing architecture across all the schemes and model optimizers was the DenseNet-161 architecture. These results indicate that the DA schemes considered in this study have demonstrated the potential to improve DL model performances. To better understand how the evaluated DL models predict clinical skin lesions, we illustrate the correlation between the actual output labels (ground truth) and the predicted outputs from the top-performing models using a confusion matrix, depicted in Figure 6. The figure demonstrates that DenseNet-161 exhibited the highest accuracy, correctly predicting around 548 images while making only 141 false predictions during the testing phase evaluation. In contrast, the other methods showed lower accuracy in correct predictions.

4.1.2. Training With Pretrained Weights

Additional experiments were carried out with the top-performing DL models (DenseNet161, ResNet101, MobileNetV3-L, and EfficientNetB7) which were trained using pretrained weights. Tables 8, 9, and 10 provide a summary of the testing and validation performance in terms of accuracy, F1-Score, and AUC, respectively. The stated DL models were trained using pretrained weights (from ImageNet) while considering different DA strategies and model optimizers for binary classification of skin lesions. The results demonstrate that models adopting DA-4 and DA-5 yielded superior performance, indicating that these augmentation strategies provided better variability and robustness during training, thereby allowing the models to generalize adequately to unseen data. The EfficientNet-B7 architecture (trained with DA-5 augmentation) attained a top accuracy and F1-Score of 85.80% in testing set evaluations. Likewise, the top AUC score was obtained with the EfficientNet-B7 architecture under similar training conditions in validation set evaluations. Experiments with DA-4 strategies also reveal superior classification performance for DenseNet161, ResNet101, and MobileNet-V3, with accuracy scores reaching 84.64%, 84.93%, and 84.93%, respectively, in testing set evaluations. Similar performance was recorded in the validation set with DenseNet161, ResNet101, MobileNet-V3, and EfficientNetB7 attaining accuracy scores of 86.38%, 86.09%, 86.38%, and 87.54%, respectively.

The choice of optimizer also significantly impacted the performances of the DL models in binary classification experiments conducted with pretrained weights. In the validation set evaluations, the EfficientNet-B7 model, trained with the Adam optimizer, attained the best AUC score of 92.34% when DA-4 strategy was considered. Conversely, in the testing set evaluations, the best AUC score of 91.02% was recorded with the DenseNet161 architecture (trained with the DA-3 strategy). While RAdam and AdamW optimizers yielded competitive performance for the models investigated, we find that when using pretrained weights, the Adam optimizer provided the most consistent results when combined with effective DA strategies.

The EfficientNet-B7 model architecture outperformed other architectures in these sets of experiments considering pretrained weights attaining binary accuracy, F1-Score, and AUC of 85.80%, 85.80%, and 90.93%, respectively, in the testing set evaluations. This architecture also demonstrates compatibility with the DA-5 augmentation strategy and the Adam optimizer. Superior performance was also recorded with the DenseNet161 and MobileNet-V3L model architectures, obtaining top accuracy of 84.64% and 84.93%, respectively, in testing set evaluations. Overall, the results obtained with pretrained models show that the combination of the Adam optimizer and DA-4/DA-5 strategies is particularly effective for the binary classification of skin lesions using clinical lesion images.

4.1.3. Statistical Analyses

A statistical comparison between the DA-1 scheme and the original dataset for DenseNet161 yielded p values of 0.95 for both the paired t-test and Welch’s t-test. We observe the statistical difference between the original (without any DA) and other DA schemes in each model architecture over different model optimizer instances. The best accuracy (79.40) was obtained with the DenseNet161 model with Adam optimizer for models trained from scratch without pretrained weights. Statistical tests were conducted using the paired t-test (12) and the Welch t-test (15) to compare the performance of the investigated models on the testing set. Conversely, with the DenseNet161 operating with the DA-3 scheme, the resulting models outperformed the case without augmentation with p values of 0.07 (paired t-test) and 0.09 (Welch t-test). The best overall accuracy with the ResNet101 model architecture was recorded with the DA-3 scheme and AdamW optimizer. Statistical comparisons of models operating DA-3 with the instance without augmentation reveal p values of 0.14 and 0.08, respectively, with the paired t-test and Welch t-test. The overall best accuracy for MobileNet-based architectures was recorded with the DA-4 scheme. Statistical comparisons here show p values of 0.007 and 0.003, respectively, in paired t-tests and Welch tests. Likewise, DA-3 (paired t-test: 0.004, Welch t-test: 0.015) and DA-5 (paired t-test: 0.014, Welch t-test: 0.017) outperformed the instance without DA for the MobileNet architecture. The DA-4 scheme produced the overall best accuracy for the Efficient-Net-B7 architecture across all the optimization algorithms. Statistical comparisons in this scenario with the instance without DA gave p values of 0.013 and 0.003 for paired and Welch t-tests, respectively. Similarly, DA-1 and DA-5 resulted in improved accuracy for EfficientNet compared with instances without data augmentation, with DA-1 yielding p values of 0.015 (paired t-test) and 0.003 (Welch’s t-test), and DA-5 yielding p values of 0.025 (paired t-test) and 0.012 (Welch’s t-test).

The DenseNet161 with pretrained weights gave an overall best accuracy of 84.64% with the DA-4 scheme and AdamW optimizer. Statistical tests measuring the performance of the pretrained DenseNet-161 across all optimizers in the DA-4 scheme against the instance without any DA gave p values of 0.404 and 0.372 for paired and Welch t-tests. Conversely, in the DA-1 scheme, which also outperformed the instance without any DA for DenseNet161, the p values obtained were 0.005 and 0.02, respectively, for paired and Welch t-tests showing significant statistical differences. The overall best accuracy (=84.93%) for the pretrained ResNet101 was obtained with the DA-4 scheme and Adam optimizer. Statistical tests comparing the ResNet101 with the DA-4 scheme and instances without DA revealed p values of 0.22 and 0.18, respectively. The MobileNetV3 architecture obtained an overall best accuracy of 84.93 with the DA-4 scheme and Adam optimizer. The statistical tests comparing the DA-4 scheme for MobileNet-V3 with the case without any DA gave p values of 0.150 and 0.211, respectively, for paired and Welch t-tests. Conversely, the MobileNet-v3 architecture also obtained accuracy values in DA1, surpassing the instance without DA. The measured p values in this instance were 0.07 and 0.09, respectively, demonstrating some statistical difference. The best overall accuracy for the pretrained EfficientNet-B7 was obtained with the DA-5 scheme. Statistical comparisons of this instance with the case without DA gave p values of 0.04 and 0.02, respectively, for paired and Welch t-tests.

4.2.1. Training With Random Weights

Table 11 summarizes the performance of the 13 different DL architectures trained on clinical lesion images to perform multiclassification tasks. All versions of DenseNet trained on original images surpass all other DL methodologies when assessed on the testing set (when attempting to solve the binary classification). They exhibit a performance ranging from 55.4% to 57.8% across the performance classification measures (accuracy, recall, precision, and F1-Score). DenseNet models achieved BACC scores ranging from 43.1% to 46.1%. Overall, DenseNet121 was the top-ranked model in three out of five performance measures, while DenseNet201 ranked second in two out of five measures. The second best performing DL architecture is ResNet50, demonstrating notably superior performance to all instances of MobileNetV3 and EfficientNet variants. Conversely, MobileNetV3 variants exhibit the worst performance levels across all the evaluation metrics. Additionally, we conducted a detailed analysis of the performance relative to each DL variant to identify the top-performing model within each architecture. Consequently, DenseNet121, ResNet50, EfficientNetB0, and MobileNetV3-S emerged as the top performers across their respective DL architectures. Therefore, we utilized the chosen DL techniques incorporating variations of the Adam optimizer (such as RAdam, Adam, and AdamW) to carry out supplementary experiments on the training set containing either original or data-augmented images. We conducted additional experiments to assess whether there is potential for enhancing DL model performance factoring DA compared to DL methods, which are trained solely on original images.

The research outcomes were consolidated to gain insights into the multiclass classification problem, considering the optimization schemes previously described, and we subsequently trained the DL methods the original and the DA images. The results of the performance metrics (accuracy, F1-Score, and BACC) obtained after training the DL techniques and then evaluating the trained models on the testing set are summarized in Tables 12, 13, and 14. From these tables, we observed that DenseNet121 factoring at least one of the optimization schemes, considering the appropriate choice of the DA scheme, obtained the highest performance, outperforming other DL methods across the observed performance measures. The DenseNet-121 architecture factoring in AdamW optimizer and trained on DA-5 yielded the highest classification accuracy and F1-Score of 65.1% and 64.3%, respectively, which outperformed the DenseNet121 model derived after training on original images (baseline) by a significant margin of 7.4%−8.4% depending on the optimization schemes (Adam or AdamW) used in the training phase. Again, we observe that The DenseNet-121 architecture factoring in AdamW optimizer and trained on DA-5 yielded the highest BACC of (55.1%) in the testing phase. To further understand the impact of the examined DA schemes, we inspect the 12 experimental scenarios: each scenario accounts (per row) for all possible experiments of a particular DL instance, factoring in one of the optimization schemes, and was trained on either original images or each of the DA schemes. Consequently, the models obtained after training were used to evaluate the testing set.

We report that the DA-5 scheme aided the DenseNet161 when trained using the AdamW optimizer and yielded the highest multiclassification metrics (accuracy, F1-Score, and BACC). Additionally, we inspected the impact of the DA across the 12 scenarios; the results show that DA-4 and DA-5 are the most influential in enhancing the DL model performance across the four main neural network architectures under study. The second competitive method is ResNet-50 architecture; when trained on DA-5 using the AdamW optimizer, it was the most favorable performance when compared with an instance of ResNet-50 trained with the other two optimization schemes (Adam and RAdam). Additionally, ResNet-50, when trained on DA-5, significantly outperforms ResNet-50 when trained on the original image by a margin ≥6.8% across accuracy, F1-Score, and AUC metrics when considering the AdamW optimizer. Likewise, the best model accuracy was recorded for the EfficientNet-B0 architecture with DA-4 compared with either Adam or RAdam optimizer: Adam performed better than the other optimization schemes. The best models from the EfficientNet-B0 architecture outperform all variants of the MobileNetV3-S in most of the performance evaluation metrics. After conducting experiments for the different instances of the MobileNetV3-S models while considering different DA schemes and optimizers, the highest model accuracy of 56.4% was achieved with DA-4 when the MobileNetV3-S model weights were optimized using the Adam optimizer. In general, the MobileNetV3-S factoring DA-4 and DA-1 schemes yielded the best performance across the three main experimental scenarios for this architecture; however, the best MobileNetV3-S trained on data-augmented images yielded the worst performance when compared with most of the best DL architectures (DenseNet121, ResNet-50, and EfficientNet-B0).

The findings demonstrate that data augmentation schemes based on spatial or geometrical transformations including random rotation (DA-4), a combination of cropping, horizontal flipping, and rotation (DA-5), horizontal flipping (DA-3), and cropping (DA-1) contribute to improved multiclass classification performance in deep learning models trained on clinical skin lesion datasets. In most instances, the performance of these DL architectures deteriorates when color jitter DA (DA2) schemes are considered. It is vital to highlight that the random rotation (DA-4) yielded the best performances in most experiments, especially with the MobileNetV3-S and EfficientNet-B0. We underscore that DenseNet121 and ResNet-50 trained with AdamW optimizer on DA-5, which factors DA-1, DA-3, and DA-4, often yielded the highest classification scores across the evaluated metrics. At the same time, the top-performing architecture across all the schemes and model optimizers was the DenseNet-121 architecture. These results indicate that the DA schemes considered in this study have demonstrated the potential to improve DL model performances. To better understand how the evaluated DL models predict clinical skin lesions, we illustrate the correlation between the actual output labels (ground truth) and the predicted outputs from the top-performing models using a confusion matrix, depicted in Figure 7. The figure demonstrates that DenseNet-121 exhibited the highest accuracy, correctly predicting around 443 images while making only 246 false predictions during the testing phase evaluation. In contrast, the other methods showed lower accuracy in correct predictions.

4.2.2. Training With Pretrained Weights

Additional experiments were carried out with the top-performing DL models (DenseNet121, ResNet50, MobileNetV3-S, and EfficientNetB0), which were trained using pretrained weights (from ImageNet) while considering different DA strategies and optimization techniques in an attempt to perform multiclassification of skin lesions. Tables 15, 16, and 17 provide a summary of the testing and validation performance in terms of accuracy, F1-Score, and AUC, respectively. The results obtained demonstrate the efficacy of models trained with DA-3, DA-4, and DA-5 strategies for improving the test and validation performances. The best multiclass classification accuracy of 75.07% was recorded with the DenseNet121 architecture utilizing the RAdam optimizer and DA-3 strategy. The EfficientNet-B0 architecture attained the highest F1-Score and BACC of 69.47% and 68.54% with the Adam optimizer and DA-4 strategy. Similar performance was observed for the EfficientNet-B0 models in the validation set with F1-Score and BACC of 72.83% and 69.83%, respectively, considering Adam optimizer and DA-4 strategy, thereby demonstrating the consistency of these models in testing and validation sets.

The choice of model optimization algorithms also affected the performances of the investigated architectures when considering pretrained weights. The Adam optimizers yielded consistent performance across different metrics with DA-4 and DA-5 strategies providing superior performance for the EfficientNet architecture in F1-Score and BACC metrics. Highly competitive performance was also observed for RAdam and AdamW in these pretrained settings. The MobileNetV3-S showed performance improvements for the BACC metric reaching about 65.15% in testing set evaluations. The DenseNet121 and EfficientNetB0 architectures outperformed others in the multiclass classification experiments involving pretrained weights. DenseNet121 attained the highest accuracy in the testing set evaluations under DA-3 and DA-5 strategies with the RAdam and Adam optimizers indicating the suitability of this architecture for multiclass classification on clinical lesion images. The deep and densely connected layers of the DenseNet model provided better feature extraction and representation in the multiclass settings.

4.2.3. Statistical Analyses

Statistical tests were conducted using the paired t-test (Eqn) and the Welch t-test (Eqn) to compare the performance of the investigated models on the testing dataset. We observe the statistical difference between the original (without any DA) and other DA schemes in each model architecture over different model optimizer instances. The best overall accuracy for the DenseNet-121 (=63.50%) trained from scratch was obtained in the DA-3 scheme with the RAdam optimizer. Statistical tests comparing the performance of the DenseNet121 in DA-3 with the instance without DA revealed p values of 0.015 and 0.0003, respectively, for paired and Welch t-tests. The ResNet-50 model trained without pretrained weights gave an overall best accuracy of 62.00% with the DA-5 scheme and AdamW optimizer. Statistical comparison of the results obtained for the ResNet-50 in DA-5 schemes with the instance without DA gave p values of 0.01 and 0.02, respectively, for paired and Welch t-tests, demonstrating high levels of statistical difference. The EfficientNet-b0 trained without pretrained weights gave an overall best accuracy of 60.3% with the DA-4 scheme and Adam optimizer. The measured p values on paired t-test and Welch tests were 0.013 and 0.012, respectively, indicating high levels of statistical difference between DA-4 and instances without DA for EfficientNet-b0. MobileNet-V3S, trained without pretrained weights, gave an overall best accuracy of 56.40 with the DA-4 scheme and Adam optimizer. The statistical tests measuring the difference between the DA-4 scheme and instances without DA for MobileNet-V3S gave p values of 0.02 and 0.0018, respectively, for paired and Welch t-tests.

The pretrained DenseNet121 model obtained an overall best accuracy of 75.07% with the DA-3 scheme and RAdam optimizer. Statistical comparison of the DA-3 scheme scenario for DenseNet121 with the instance without DA gave p values of 0.19 and 0.09, respectively, for paired and Welch t-tests. The overall best accuracy (=72.46%) obtained for the pretrained ResNet-50 architecture was obtained with the DA-5 scheme and the RAdam optimizer. Statistical comparison of the DA-5 scenario with the instance without DA for ResNet-50 gave p values of 0.11 for both paired and Welch t-tests. DA-4 also produced better performance for the pretrained ResNet50 architecture and gave p values of 0.007 and 0.008 compared to the instances without DA, indicating a high statistical difference. MobileNet-V3S trained using pretrained weights recorded the best accuracy values across all optimizers with the DA-4 scheme. Statistical comparisons of MobileNet-V3S on DA-4 with the instance without DA gave p values of 0.009 and 0.005, respectively, showing high statistical differences. The pretrained EfficientNet-b0 obtained an overall best accuracy of 73.04 in DA-4 and DA-5 schemes. Statistical comparison for EfficientNet-b0 models in DA-4 scheme with the instance without DA gave p values of 0.002 and 0.05, respectively, for paired and Welch t-tests, indicating significant levels of statistical difference.

5.1.1. SHAP Assessment on Binary and Multiclass Tasks

In this section, we provide detailed explainability discussions using feature attributions from the SHAP algorithm. The discussion in this section is premised on the DenseNet161 model trained with the DA-4 strategy for binary classification and DenseNet121 model trained with the DA-5 strategy for multiclass classification. The SHAP plots are produced for these models and trained from scratch using pretrained weights with Adam, AdamW, and RAdam optimizers. The models are analyzed using instances where correct and incorrect predictions were obtained. Figures 8, 9, and 10 provide the binary classification SHAP visualizations of the feature attributions for the DenseNet161 in pretrained and random weight settings using Adam, AdamW, and RAdam optimizers, respectively. In Figure 8, for the correct predictions, Adam-optimized models focus on the irregular borders and pigmentation, as highlighted by the red regions, for malignant predictions. For benign predictions, the model highlights the smoother areas characterized by uniform texture and coloration. Wrong predictions were obtained when the model incorrectly highlighted noncritical regions of images and areas, particularly displaying malignant features. The AdamW-optimized model highlights lesion border areas and texture irregularities as critical features for malignancy as shown in Figure 9. The lack of color variations and uniform textures was highlighted for correct benign predictions in the AdamW-optimized model. For incorrect predictions, the model attributes less significant regions and minor artifacts, failing to give significant weights to malignant features. Furthermore, it appears the model identifies elevation as a predominant feature for obtaining malignant predictions, leading to misclassifications in Figures 9(b) and 9(d). In Figure 10, the RAdam-optimized model highlights the lesion edges and heterogeneity for distinguishing malignant and benign lesions. Although the model highlights the lesion areas in the incorrect predictions, it fails to interpret the features correctly to obtain either malignant or benign predictions.

Figures 11, 12, and 13 provide the multiclass classification SHAP visualizations of the feature attributions for the DenseNet161 in pretrained and random weight settings using Adam, AdamW, and RAdam optimizers, respectively. In Figure 11, the Adam-optimized DenseNet-121 with pretrained weights focuses on the darker and irregular regions for correct BCC predictions. Uniform pigmentation and well-defined borders were highlighted as attributions for NEV prediction, while melanoma attributes were due to uneven colors and irregular borders. Incorrect predictions of SCC were due to the attribution of surface irregularities and color variations shared with BCC and ACK conditions. The AdamW-optimized model in Figure 12 highlights malignant features showing irregular pigmentation and border irregularities to obtain accurate predictions for MEL conditions. Characteristic features showing raised edges and pearly colors were attributed to BCC predictions. Incorrect predictions in this case were primarily due to the model’s feature attributions that were not definitive for the correct class. Accurate predictions for MEL conditions in Figure 13 for the RAdam-optimized model were due to features distinctive of asymmetry and color variations. Likewise, the model focuses on distinctive features with irregular textures and pigmentations for ACK and BCC. Incorrect predictions, especially in the case of NEV misclassified as MEL, were due to emphasis on features that were not indicative of malignant lesions.

5.1.2. CAM Interpretability Assessment on a Multiclassification Task

The operational functionality of our developed system is showcased in Figure 14. This explainable DL system demonstrates the capability to predict skin lesions, suggest the specific type of skin lesion (malignant or benign), and pinpoint the precise region of interest containing the lesion (tumor). Suppose a user of the developed interpretable DL platform chooses the multiclass skin lesion predictive system menu tab; upon uploading the skin image and selecting DenseNet121 and then clicking on the Predict button, the user activates the DL engine to classify the skin image to determine the target class from the six possible class categories (displaying the extent of certainty). The result demonstrates the system effectively predicted the provided skin image containing BCC, a type of malignant tumor, revealing a certainty score of 100%. Conversely, the confidence scores for other classes (SEK, NEV, ACK, SCC, and MEL) were 0%. Consequently, we chose the interpretable scheme (GradCAM) from the available options and then clicked on the interpret button. The interpretability analysis showed that the BCC can be found in the middle portion of the clinical skin image. We conducted another assessment of DenseNet121 using the other interpretable algorithms (AblationCAM and GradCAM++). Figure 15 demonstrates that all interpretable algorithms concurred and identified nearly identical attention regions (localized portions) in the skin containing BCC. This comprehensive analysis is crucial for medical experts to establish confidence before proffering treatment protocols.

5.1.3. CAM Interpretability Assessment on a Binary Classification Task

We conducted a detailed examination of the operational features within our developed system’s binary skin lesion predictive system menu, as depicted in Figure 16. Suppose a user of the developed interpretable DL platform chooses the binary skin lesion predictive system menu tab; upon uploading the skin image and selecting MobileNet-Large (MobileNetV3-L) and then clicking on the Predict button, the user activates the DL engine to classify the skin image to determine the target class from the two possible class categories (displaying the extent of certainty). The result demonstrates the system accurately predicted the provided skin image containing a benign tumor, achieving a confidence score of 100%. Conversely, the confidence score for a malignant tumor is 0%. Furthermore, the explainability analysis suggests that the benign tumor is located at the upper left and lower middle regions of the clinical skin image. We replicated the assessment using the other interpretable algorithms and present the resulting observations in Figure 17. The figure shows that all the interpretable algorithms agreed and identified nearly identical attention regions (localized portions) in the skin containing benign tumors. Consequently, this attention region influenced MobileNetV3-L’s decision to classify the uploaded skin image as depicting a benign tumor. Such thorough analysis is essential for medical experts to devise treatment protocols confidently.