ENeTAMIB: an attention-inception-based model for accurate and explainable skin disease diagnosis

Objectives

This research rapidly detects skin cancer using images captured with a dermatoscope. Early detection of skin cancer by physicians enhances the likelihood of timely intervention and halting disease progression. It also evaluates the efficacy of proposed model in comparison to the most advanced medical deep learning models. This research classifies various forms of skin cancer using dermoscopic images.

• To develop and implement baseline model using two inception and inception modules for multi-class skin disease classification based on dermoscopic images. To enhance model performance by integrating architectural components, and transfer learning for better feature extraction and class discrimination.

• To address the limited and imbalance dataset problems, we use augmentation and synthetic minority overampling technique.

• To rigorously evaluate the model using k-fold cross-validation, performance metrics (accuracy, precision, recall, F1-score), and statistical significance testing (paired t-tests) to ensure robustness and meaningful improvement over baseline models.

• To improve model transparency and clinical trust by applying explainable artificial intelligence (XAI) techniques such as gradient weighted class activation mapping (Grad-CAM) and local interpretable model agnostic explanations (LIME), and by testing the model on an external dataset (HAM10000) to assess generalizability.

Contributions

The study has following main contributions:

• This research developed a novel architecture based on an attention module and inception blocks that works well for detecting 57 skin diseases. The main structure of the model is EfficientNet-B2, and it has an inception module, an attention module, global average pooling, and fully connected layers, which improved feature identification and led to better accuracy in classification compared to basic models like Visual Geometry Group-16 (VGG16) and Residual Neural Network 50 (ResNet50).

• For evaluating the proposed model, the augmentation approaches improved the skin samples and prevented overfitting. After augmentation, the performance improved by a large margin. The balanced data provides CI [0.9822–0.9905] and 0.1041 validation loss.

• XAI techniques such as Grad-CAM, saliency maps, and LIME provide the model’s decision-making capabilities. The regions of the image that are crucial to the model’s evaluation are highlighted in a heat map produced by GRAD-CAM, which shows their ability to generalise and provide visual justification for predictions. Saliency maps provide more accurate possibilities and are more complicated, which may significantly affect the design.

• The proposed model is compared with the most recent state-of-the-art studies in the literature, primarily from the last 2 years, 2023 and 2024. In addition, five-fold cross-validation and statistical analysis (paired t-tests) demonstrate consistent performance and statistically significant improvements.

The remainder of the article is structured as follows: The literature review is covered in ‘Literature Review’, the proposed methodology is demonstrated in ‘Materials and Methods’, the experimental results are explained in ‘Results and Discussion’, and the conclusion is presented in ‘Conclusion and Future Directions’.

Dataset description

This study used the Skin Diseases and Cancer dataset, freely available at Mendeley (Mafi et al., 2023). The term “skin diseases” refers to a wide range of conditions impacting the biggest organ in the body, from everyday skin disorders to more serious and even deadly diseases. Dermatological disorders, such as sun damage, inherited traits, or environmental factors, may cause the excessive growth of skin cells. Melanoma, basal cell carcinoma, and squamous cell carcinoma are major threats to world health because of their frequency and metastatic potential. But millions of people deal with harmless skin problems like acne, eczema, and psoriasis, which may harm their quality of life and, if left untreated, can cause serious complications. This collection has 978 unique images of skin diseases and cancer, including 90 from primary sources and 888 from secondary sources.

Skin image preprocessing

Image preprocessing entails converting raw image data into a format that is both comprehensible and operational. It facilitates the improvement of essential attributes vital for computer vision applications while removing undesirable distortions. The first stage of preprocessing is crucial before supplying data to deep learning models. The optimal performance of deep learning algorithms relies on the consistent scale of images (Zghal & Derbel, 2020). To modify the proportions of images, we use the resize() function in OpenCV. We use a fixed size (150 width, 150 height, and three channels) for the experiments.

Skin image augmentation

Augmentation is a collection of strategies that can increase the volume of available data in current datasets. Augmentation is a beneficial way to shrink deep learning models, which might help prevent overfitting or improve performance when the initial dataset is not enough for training. An individual’s capacity to prevent overfitting is only one of the many advantages of augmentation (Han et al., 2020). A substantial sample is essential for the effective functioning of machine learning and deep learning models. To enhance the efficacy of the model, we augment its current data set. This statement essentially suggests that modifications might enhance the model’s performance. Collecting data and categorizing it for machine learning models may be costly and labor-intensive. Medical professionals can identify cost-saving opportunities by improving information through data enrichment initiatives. Skin disease samples taken from the augmented dataset are shown in Fig. 2. The total number of samples before and after using the augmentation is presented in Table 1. The techniques are shown below:

• Image flipping is the simplest kind of augmentation technique. As the name implies, it creates a new version of the image by flipping it. has a horizontal image reflection. Conversely, it flips images horizontally from right to left. It can be oriented either horizontally or vertically.

• A common technique for data augmentation is image rotation. We may broaden the variety of our training data and strengthen our model by changing the viewpoints of the images.

• We use image cropping to emphasize specific areas of an image. This technique is especially effective in scenarios like item identification when the objective is to have the model identify things independently of their location within the image.

• The horizontal flip function can flip an image horizontally, while the vertical flip feature can flip it vertically. Most augmentations allow for the parameter p, which controls the probability of using the augmentation.

• The dimensions of the image may be altered by the scaling technique. Image shear involves relocating a segment of an image while keeping the other portions intact.

Skin data splitting

In data splitting, we utilized train test split and cross validation two stage approaches to ensure the model robustness and reduced the overfitting risks. Initially, the entire dataset was divided into two subsets: 70% for training and 30% for testing. This set of combined tests was used to evaluate the final model to ensure that performance metrics reflected the model’s ability to reproduce untrained data. On the dataset, we performed five-fold cross-validation during training to tune the model parameters and improve the overall performance. During the fold cross-validation, the training dataset was divided into five different subsets. Each subset serves as a validation vector, and the remaining four subsets are used for training. This process was repeated five times, and the average of the five iterations provided the most accurate and reliable estimate of model performance.

Methodology

This work used EfficientNet-B2 as the backbone architecture, using transfer learning from ImageNet weights, deactivating the top layer of EfficientNet, and selecting a fixed input skin image size for the model. Transfer learning enables the model to operate more efficiently and provide superior and more precise results. During the setup, we suspend training on some levels of the backbone model after the first training session. This facilitates the identification of distinctive patterns within the data. The proposed ENeTAMIB architecture for the classification and detection of 57 skin diseases is illustrated in Fig. 3.

In step 2, we construct an attention module including a global average pooling layer, a dense layer using ReLU activation, and a subsequent dense layer employing a sigmoid activation function. Subsequently, multiply them by the function variable. In the third step, we develop the Inception module function to integrate into the backbone architecture. Initially, we configure a Conv2D layer using a 1 by 1 filter with ReLU activation, followed by two further Conv2D layers, filter [1], (1,1) and filter [2], (3,3) again employing ReLU activation. Subsequently, we include a max-pooling layer with a stride of 1 by 1. The last layer in the Inception module is once again a Conv2D layer using ReLU activation. Variables 1, 2, and 3 are concatenated and returned to the function. Following the incorporation of attention and inception modules into the backbone architecture, we implement two-dimensional global average pooling inside the architecture. To enhance model efficiency, a fully linked layer and a ReLU activation layer are used with dropout. We then use the final classification layer, Softmax, for 57 skin conditions.

We employed Inception and Attention modules with the EfficientNetB2 backbone, to improve the feature representation capability of the proposed model. Inception and Attention components serve complementary roles in improving classification performance for multi-class skin diseases as represented in Algorithm 1.

The Inception module is designed to capture multi-scale spatial information by processing the same input through multiple parallel convolutional paths with different kernel sizes. In this work, three branches operate simultaneously: (1) a 1 $\times$ 1 convolution for dimensionality reduction, (2) a 1 $\times$ 1 convolution followed by a 3 $\times$ 3 convolution to extract mid-level spatial features, and (3) a 3 $\times$ 3 max pooling layer followed by a 1 $\times$ 1 convolution to retain contextual information while reducing spatial resolution. The outputs from all branches are concatenated along the channel dimension, allowing the network to learn features at multiple scales efficiently.

In parallel, the Attention module is introduced to help the network focus on the most informative channels. Specifically, we use a channel-wise attention mechanism inspired by the squeeze-and-excitation block. To summarize local features in channel analysis starts from the global average level. The feature passes through two fully connected layers with a threshold matrix to learn the nonlinear structure of the channel. The designed prediction weights are applied to the original feature maps by adding parameters, each weight is re-evaluated based on its learned value.

Activation functions are essential for neurones in a neural network to process inputs and produce outputs. Among the several activation functions used by contemporary deep learning models, the rectified linear unit (ReLU) is particularly notable. It excels in many neural network topologies and is user-friendly, which mostly accounts for its popularity. ReLU outputs the whole input if it is positive; else, it outputs zero. This characteristic not only preserves the non-linearity essential for comprehending intricate patterns but also mitigates issues such as the vanishing gradient problem that may occur in deeper networks. Due to its efficiency in facilitating model learning, ReLU is a favoured choice of this study.

In the proposed attention module, we use the sigmoid function. The sigmoid function was one of the earliest activation functions introduced to neural networks. The sigmoid’s smoothness and differentiability make it an excellent choice for backpropagation. Because its output may range from 0 to 1, it is quite helpful for the classification challenges. In the proposed model, the layers which are primarily used for image classification tasks such as disease diagnosis, the transition from convolutional feature extraction to final classification requires flattening or pooling of the feature maps. The global average pooling2d layer serves this purpose by computing the average value of each feature map like channel across its spatial dimensions. We did not used here traditional flattening,we used global average pooling that reduces each feature map to a single value. This approach provides better accuracy, reduces overhead, and increases the overall cost of maintaining the integrity of the training data. After the transformation, the network uses a convolutional layer of 64 neurons as the ReLU activation layer. This step accounts for the nonlinearity and allows the model to learn the baseline signals from the interacting variables. Finally, a dropout of 0.5 was used to remove half of the neurons in each training iteration. The regularization allows the network to learn a variety of features to avoid misclassification of any neural network. These features act as a transformation of the spatial distribution towards the population, resulting in a nonlinear representation and a high degree of variability in the dataset. A fundamental cost function is cross-entropy loss. It’s principal function is to facilitate the optimisation of classification models. The model includes a softmax activation function and a cross-entropy loss function. Employing this loss to train the proposed model enables the generation of a probability for each image over all 57 classes, facilitating classification into multiple categories.

Overall performance of proposed ENeTAMIB model

The performance of the proposed ENeTAMIB model using the original 5,000 images and 10,000 images is presented in Table 3. The results of the proposed model for the last ten epochs, along with training and testing accuracy, are presented. We fit the proposed models for 100 epochs to achieve optimal results and enhance model training. The model was learnt efficiently for 100 epochs.

We also calculate the average of the last ten epochs, standard deviation, entropy, and mean absolute deviation for the proposed model. At last epochs, training accuracy using original data is 0.9855, and testing accuracy is 0.4157. The sample size for both sets is extremely small; the model typically receives one image per class for testing purposes. That was very low; that’s why the model achieved very low performance on original data. To overcome this data limitation, we used augmentation to enhance the dataset. For 5,000 enhanced images, the model achieved 0.9900 training and 0.9407 testing accuracy, which is somehow better than 0.4157. Again, we did experiments with 10,000 enhanced images to see the performance of the proposed model. The model achieved 0.9933 training and 0.9863 testing accuracy. We observed that the suggested model excels in handling a substantial image data set, as deep learning models excel in handling large datasets.

Learning curves of multi-class proposed ENeTAMIB model

An important technique in artificial intelligence, a “learning curve” depicts the correlation between a model’s performance and its training data or the progression of validation. Achieving this outcome requires only monitoring the model’s accuracy throughout training and validation while including additional training examples. All learning curves contain these essential elements: The training curve serves as an indication of the model’s performance throughout its learning process on the training dataset. The model initially exhibits superior performance due to the extensive data included in the training set. The validation curve illustrates the model’s efficacy on a separate validation dataset after training is completed. This feature enables us to evaluate the model using new, untested data. Figure 4A illustrates the accuracy plots for the proposed model using the original data. A significant disparity exists between the training and validation curves. Figure 4B illustrates the correctness of the proposed model with 5,000 improved images, demonstrating that the model has learnt perfectly. In Fig. 4C, it is evident that the curves first decline, followed by improvements after several epochs. Figure 4D illustrates the loss plots for the proposed model using the original data. A significant disparity exists between the training and validation curves. Figure 4E illustrates the correctness of the proposed model with 5,000 improved images, demonstrating that the model has learned perfectly. In Fig. 4F, it is evident that the curves first decline, followed by improvements after several epochs. The model achieved very low validation loss with the augmentation technique.

Explainability using Grad-CAM visualization for the proposed ENeTAMIB model

Class activation mapping (CAM) enables the identification of certain regions in an image that are essential for addressing a particular classification challenge. Since its release by Zhou et al. (2016), it has rapidly gained popularity as a technique for interpreting predictions generated by deep learning convolutions. Initiatives are underway to elucidate and enhance comprehension of deep learning. Enhancing the interpretability of models is essential for several deep learning applications in medical imaging. The Grad-CAM approach is beneficial for analysing closely linked neural networks in detection or prediction tasks by offering both model insights and visual explanations. GradCAM’s comprehensive generalisation of CAM makes it superior to CAM.

Grad-CAM and LIME visualizations are included in this work because these play an essential role in bridging the gap between high-performing models and their adoption in clinical practice. However, their contributions must be critically evaluated beyond surface-level interpretability. LIME provides locally faithful explanations by approximating complex model behavior with interpretable surrogate models around individual predictions. This allows clinicians to understand which image regions influenced a prediction. However, LIME’s perturbation-based approach is inherently unstable: explanations can vary significantly across runs due to randomness in the sampling process. Moreover, it is model-agnostic and lacks a deep understanding of the image’s spatial structure, which is critical in dermatology. It may limits its reliability in high stakes diagnostic decisions where consistency and reproducibility are more essential.

On the other hand, Grad-CAM generates class-discriminative heatmaps by leveraging the gradients flowing into the last convolutional layers. It provides visual insight into what the model “sees” as important. While Grad-CAM is more stable and spatially coherent than LIME, it too has limitations. It is sensitive to architecture design and may produce misleading saliency in some cases. Moreover, it only explains where the model looked, not why a decision was made.

From a clinical perspective, while both methods enhance model transparency and offer visual clues that can build initial trust, they are not sufficient for clinical-grade reliability. True deployment requires further validation, integration with domain knowledge, and consistent interpret-ability across diverse populations and image qualities.

Grad-CAM visualization of 3rd last convolution layer for the proposed model is shown in Fig. 5A. GradCAM’s comprehensive generalisation of CAM makes it superior to CAM. Grad-CAM visualization of 2nd last convolution layer for the proposed model is shown in Fig. 5B. GradCAM’s comprehensive generalisation of CAM makes it superior to CAM. Grad-CAM visualization of last convolution layer for the proposed model is shown in Fig. 5C.

Explainability using saliency maps

Numerous methods exist to correlate a network’s predictions with its input attributes in supervised learning. The saliency map is constituted by the variance of the class score across the input pictures. By instructing a machine learning model to analyses an image and identify its most significant components, one may generate a saliency map. The objective is to identify the pixels on which the model concentrates during prediction. This will provide insights into the model’s decision-making processes and its internal representations. To enhance the accuracy of a saliency map for a certain image, it samples analogous images, introduces noise, and then averages the saliency of the sampled images. The explainable proposed model visualization with the saliency map is represented in Fig. 5D.

The use of confusion matrix findings is a straightforward approach that proves to be very efficient when assessing the performance of a model on two or more variables, as seen in Fig. 6 for 57 variables. In a style similar to a table, it presents the actual results on one side and predictions on the other side. TP, TN, FP, and FN are all four aspects that come together to form the matrix. It is possible to have a thorough understanding of the model’s strong and weak areas by using these classifications.

Results visualization of LIME, skin images in yellow indicate super-pixel segmentation and in blue indicate LIME heat-maps are presented in Fig. 7. A notable benefit of LIME is its simplicity. The LIME system is exceedingly intricate; however, its core principle remains straightforward and transparent. LIME seeks to clarify the behaviours predicted by the model in certain situations. Even the most basic models may effectively forecast their behaviour by focussing on a sufficiently limited decision surface.

Multi-class classification performance of proposed ENeTAMIB model

The multi-class classification performance of the proposed ENeTAMIB model for 57 skin diseases is shown in Table 4. The proposed model attained a macro precision average of 0.9851, a macro recall average of 0.9891, and a macro F1-score average of 0.9869. Class label 0 attained a perfect 100% across every metric, whereas label 10 recorded a precision of 0.9881, a recall of 0.9765, and an F1-score of 0.9822. Label 20 attained an accuracy of 0.9811, a recall of 0.9630, and an F1-score of 0.9720. Label 43 got a perfect score of 100%.

Comparison of proposed model with previous studies

This section analyses the proposed model with a literature review and contrasts the experimental results of our proposed model with those of the most existing models. In Table 5, a thorough comparison of the proposed model and the state-of-the-art models is conducted. We made comparison of current model with latest 2023 and 2024 literature cited in the study. We seen that three researchers in the literature used CNN model, and attained more than 95% accuracy without applying augmentation technique. Arshed et al. (2023) employed vision transformer for seven skin cancer and attained only 92.14% accuracy. Ayas (2023) also utilized swin like transformer and attained unsatisfactory performance. Ahmad et al. (2023) utilized shuffleNet model with augmeneted large number of images (36,000) for the experimentation and attained 91.5% accuracy. ShuffleNet after enhanced data, also got less accuracy. Another work (Radhika & Chandana, 2023) employed the Multi-class Skin Cancer Detection Network (MSCD-Net) for 25,331 skin images and attained 98.77% accuracy using four diseases. They got higher accuracy because the dataset has only four disease to for the classification. Alhudhaif et al. (2023) and Khan et al. (2024) used convolution support vector machine and in the article the number of images for the experimentation was not mentioned. They attained 89% accuracy via data augmentation for only seven disorders. All research referenced in Table 5 exhibited poor accuracy for multi-class classification, and some studies did not augment the dataset to increase the results. The proposed model has shown excellent performance in detecting 57 skin diseases and also offered explanations to elucidate its predictions. Here, we also performed experiments using the balanced training data. After applying Synthetic Minority Over-sampling Technique (SMOTE) to the training data, we achieved the same number of samples for each class. The proposed model achieved 98.63% accuracy and 0.1041 validation using balanced data.

Ablation study

We performed an ablation study because it is essential to validate the individual contribution and efficacy of the proposed model. This provides empirical evidence to support our design choices and performance claims, ensuring that improvements are not due to incidental factors. Here, in Table 6, the baseline EfficientNet+Inception+Attention framework achieved 97.07% accuracy with 0.1272 validation loss. When we removed the attention layer, the model achieved 95.10% accuracy and 0.1764 loss. Again, we also removed the Inception layer and achieved 94.43% accuracy and high validation loss. After, we fine-tuned the EfficientNet+Inception+Attention framework and fit it with only 20 epochs and achieved 98.07 accuracy and low validation, validating that the designed model is superior.

Cross validation study

Cross-validation is a technique for evaluating models that entails their training on subsets of the available input data and their subsequent evaluation on the complementary subset of data. Cross-validation is employed to detect overfitting, which is the inability to generalize a pattern. Five-fold cross-validation was implemented. Table 7 presents the results using five-fold cross validation.

Statistical tests

The paired statistical t-test is used to determine whether the performance difference between models is statistically significant or simply due to chance. The test produces, P-value if this value is less than 0.05, the difference is considered statistically significant at the 95% confidence level. This means that there is strong evidence that one model is superior to the other. If the P-value is greater than 0.05, it indicates that the difference in performance is not random and is not significant. Results for statistical tests and P values are shown in Table 8. We performed tests with the VGG16 and ResNet50 baselines with the proposed model and achieved significant results.

Confidence interval

A confidence interval (CI) provides a statistical range that estimates how reliable a performance metric such as accuracy is. The proposed model achieves 98.60% accuracy with a CI of [0.9818–0.9902]; as shown in Table 9, it means we are 95% confident that the true accuracy on unseen data falls within that range. CI is important in clinical settings, where consistent and dependable performance is critical. CI helps quantify uncertainty due to sample size or data variability and supports better decision-making when comparing models. The VGG16 base model achieved 73.93% accuracy with a CI of [0.7236–0.7550], and ResNet50 achieved 91.27% accuracy with a CI of [0.9026–0.9228]. In Eq. (5), $\widehat{oa}$ represents the observed accuracy while z represents the standard value.

(5) $${\mathrm{C}\mathrm{I}}_{0.95}=\hat{o}a\pm z\cdot \sqrt{\frac{\hat{o}a(1-\hat{o}a)}{n}}.$$

Experiments using balanced training data

Table 10 presents the results using balanced training data. We used SMOTE to balance the 57 skin diseases, and after SMOTE, we achieved the same number of images for each disease type for experimentation. We used the SMOTE only on training data, and test data remained unchanged. We used 100 epochs for model fitting with a batch size of 64. Here we show the results for the last 10 epochs, from the 91st to the 100th epochs, and mention them in the table from 1 to 10. The proposed model achieved 98.59% average accuracy with 0.0021 STD.

Experiments using the HAM10000 dataset

We also performed experiments using HAM10000 dataset, and the dataset was splitted using 70:30 ratio and fitted with the 20 epochs. It corresponds to a large collection of multi-source dermatoscopic images of common pigmented skin lesions. The dataset has actinic keratoses and intraepithelial carcinoma (akiec), basal cell carcinoma (bcc), benign keratosis-like lesions (bkl), dermatofibroma (df), melanoma (mel), melanocytic nevus (nv), Vascular lesions (vasc). Table 11 shows the results of the model using HAM10000 dataset.