An Explainable Computer-Aided Framework for Skin Lesion Classification Using Deep Learning

Summary

This paper presents an explainable CNN-based framework for skin lesion classification that combines high diagnostic accuracy with model interpretability. It classifies lesion images and generates visual explanations for its predictions. Results demonstrate strong performance and improved transparency, supporting clinical decision-making and assisting dermatologists in the early detection of skin diseases.

Abstract

The utilization of deep convolutional neural networks for the purpose of diagnosing diseases in the skin area has proven to yield similar accuracy levels to those obtained by dermatologists in different studies. Nevertheless, many challenges are still present, including underperformance and poor generalization in some cases, as well as low interpretability related to the use of black box models. This creates major obstacles for practical implementation since it requires explainability in addition to accurate diagnostics, making it necessary to find a solution. To overcome the mentioned difficulties, an explainable deep learning framework for skin lesion classification (EDLF-SLC) is developed within this study. The framework makes use of various approaches in machine learning and explainable artificial intelligence (XAI) to ensure improvements in terms of both accuracy and interpretability in a three-stage manner. In the initial stage, deep representations extracted from multiple pretrained CNN architectures are fused to preserve complementary discriminative information learned by different network architectures before classification using an SVM with a radial basis function kernel. Next, Support Vector Machine (SVM) classifiers with a radial basis function kernel are used to classify the obtained features. Finally, the predictions made by the model are interpreted through local interpretable model-agnostic explanations (LIME). Experimental results show that EDLF-SLC reaches 88.0% in accuracy, 89.0% in precision, 87.0% in recall, and 88.0% in F1 score, demonstrating competitive performance compared with several recently reported methods under the experimental conditions considered in this study.

Introduction

Skin lesions are a major concern in dermatology and oncology because they range from benign abnormalities to malignant cancers such as melanoma, the deadliest form of skin cancer. In 2020, melanoma accounted for approximately 325,000 new cases and more than 57,000 deaths worldwide¹. Although advances in screening and treatment have improved patient outcomes, delayed diagnosis and limited access to healthcare continue to contribute to high mortality and years of life lost, particularly among younger populations^2,3.

Traditional diagnosis relies primarily on visual examination and dermoscopy, making the process dependent on clinicians' expertise and prone to subjectivity, inter-observer variability, unnecessary biopsies, and prolonged examination times^4,5,6. To address these limitations, deep learning, particularly CNNs, has emerged as an effective solution for automated skin lesion classification. CNN-based models, including DDCNN-F⁷, YOLOv7-XAI⁸, and several other architectures^{9,10,11,12,13}, have demonstrated high diagnostic accuracy by automatically learning discriminative image features. Despite their success, most deep learning models operate as "black boxes," limiting clinicians' trust and hindering their adoption in routine medical practice. Explainable artificial intelligence techniques have therefore been introduced to improve model transparency by providing visual explanations of predictions. Among these methods, Local Interpretable Model-Agnostic Explanations generates interpretable local explanations by identifying the image regions that most influence model decisions¹⁴.

Skin lesion classification has evolved from traditional clinical assessment to advanced AI-based diagnostic systems driven by the need for accurate, reliable, and early skin cancer detection. This evolution has progressed through five major stages—traditional diagnostic methods, computer-assisted diagnosis (CAD), ML, DL, and, more recently, XAI and federated learning (FL)—reflecting continuous efforts to improve diagnostic accuracy, consistency, scalability, and clinical trustworthiness while overcoming the limitations of conventional examination. Early computational methods attempted to emulate clinical reasoning through handcrafted image descriptors derived from the ABCD rule, including asymmetry, border irregularity, color variation, and lesion diameter. These features were combined with Bayesian networks, decision trees, expert systems, and fuzzy inference models to support melanoma diagnosis, with several studies reporting classification accuracies exceeding 90%^15,16,17. Despite providing an important foundation for computer-aided diagnosis, these methods depended entirely on manually engineered features and predefined diagnostic rules. Consequently, they exhibited limited robustness against variations in image quality, lesion morphology, illumination, and acquisition conditions.

To address these shortcomings, classical CAD systems emerged as an intermediate step between conventional image analysis and modern AI-based frameworks. CAD systems combined handcrafted feature extraction with conventional classifiers to improve diagnostic consistency and assist clinical decision-making. Several hybrid approaches integrated hierarchical clustering with recurrent neural networks and long short-term memory architectures, achieving classification accuracies approaching 99.5%¹⁸. Other CAD-based frameworks incorporated gradient boosting classifiers with SHAP-based explanations, demonstrating competitive diagnostic accuracy while improving model transparency¹⁹. Although these systems represented a significant improvement over purely rule-based approaches, they continued to rely heavily on handcrafted feature extraction, extensive preprocessing, carefully annotated datasets, and multistage processing pipelines.

Machine learning techniques further advanced automated skin lesion classification by enabling data-driven learning rather than explicit rule design. Hybrid frameworks combining convolutional neural networks with conventional machine learning classifiers, including logistic regression and k-nearest neighbors, demonstrated strong classification performance on benchmark datasets, achieving accuracies exceeding 95%⁹. Self-supervised multimodal learning further improved feature representation by jointly exploiting dermoscopic and clinical images, thereby reducing dependence on extensive manual annotation while improving classification performance²⁰. Additional studies combined CNNs with generalized discriminant analysis, SURF descriptors, and optimization algorithms to improve feature discrimination and classification accuracy²¹. Automatic feature selection techniques employing DenseNet-201, InceptionV3, and binary tree optimization algorithms further enhanced feature representation while maintaining competitive diagnostic performance²². Transfer learning approaches utilizing pretrained architectures such as AlexNet and GoogLeNet also demonstrated promising classification capability despite limited training data²³. Nevertheless, most ML methods continued to depend on carefully designed preprocessing procedures, handcrafted optimization strategies, balanced datasets, and extensive hyperparameter tuning. Their limited external validation and sensitivity to class imbalance further constrained their practical applicability in diverse clinical settings.

The introduction of DL fundamentally transformed automated skin lesion classification by enabling end-to-end learning directly from raw image data. CNNs eliminated the need for handcrafted feature engineering while substantially improving diagnostic performance across multiple benchmark datasets. Most CNN-based approaches demonstrated significant improvements in lesion classification through increasingly sophisticated architectures. Symmetry-aware CNN models explored clinically meaningful lesion characteristics but achieved only moderate balanced accuracy²⁴. Others integrated EfficientNet architectures with LIME and SHAP explanations to improve interpretability while introducing quantitative trustworthiness measures²⁵. Hybrid DL systems combining lesion segmentation, ensemble learning, and CNNs achieved excellent performance on multiple ISIC datasets but remained sensitive to segmentation quality and class imbalance²⁶.

More recently, increasingly sophisticated hybrid architectures have further improved classification accuracy by integrating complementary DL techniques. Conditional generative adversarial networks combined with YOLOv5 and independent component analysis achieved classification accuracies exceeding 99%, although their computational complexity limited practical deployment ²⁷. Similarly, hybrid LSTM–ResNet architectures effectively learned spatial-temporal representations but required substantially greater computational resources²⁸. Transformer-based models, including Sap-Former, exploited global contextual information to enhance feature representation but required extensive preprocessing, large-scale annotated datasets, and significant computational power²⁹. Lightweight alternatives such as S-MobileNet attempted to balance computational efficiency with diagnostic accuracy³⁰, whereas unsupervised domain adaptation methods improved cross-domain generalization despite reduced effectiveness when only limited training samples were available³¹. Attention-enhanced hybrid networks incorporating modified convolutional attention modules and snapshot ensemble learning also demonstrated promising performance for monkeypox skin lesion classification, although challenges associated with class imbalance, image diversity, and limited explainability remained unresolved³². Customized residual architectures employing deeper network designs and hybrid loss functions further improved classification accuracy beyond 99% but incurred considerably greater computational overhead and training time³³. Comprehensive review studies consistently concluded that DL substantially outperforms traditional handcrafted approaches by automatically learning discriminative image representations, while simultaneously identifying persistent challenges associated with imaging artifacts, illumination variability, computational complexity, and limited clinical generalization³⁴.

Although DL has dramatically improved diagnostic accuracy, concerns regarding model transparency, uncertainty estimation, and trustworthy clinical deployment have stimulated growing interest in XAI and federated learning. Several studies have investigated uncertainty quantification techniques, including conformal prediction, Monte Carlo dropout, and evidential deep learning, demonstrating that reliable confidence estimation can substantially improve decision support despite imposing additional computational and data-related constraints³⁵. Transformer-based mutual learning frameworks have also been proposed to address class imbalance while improving feature learning efficiency³⁶. Other approaches combined full-resolution CNNs with graph-based optimization techniques to improve lesion segmentation and classification³⁷, whereas hybrid deep learning frameworks integrating multiple complementary feature representations achieved classification accuracies approaching 99.5% despite requiring extensive preprocessing³⁸.

Recent research has increasingly focused on integrating explainability directly into DL frameworks to enhance clinician trust and facilitate practical clinical adoption. Explainable systems employing multiple convolutional neural network architectures together with Grad-CAM and Score-CAM successfully localized diagnostically relevant lesion regions while maintaining competitive classification performance on both internal and external datasets³⁹. Hybrid CNN–Transformer frameworks combining dermoscopic images with clinical metadata further improved predictive performance while utilizing SHAP and Grad-CAM to generate clinically meaningful visual explanations⁴⁰. Additional transformer-based hybrid architectures integrating EfficientNetV2S, Swin Transformers, modified Xception networks, and graph attention mechanisms effectively captured complementary global and local lesion characteristics, although their large parameter counts and limited performance on minority classes restricted practical deployment⁴¹. Similarly, hybrid YOLOv8–Vision Transformer frameworks demonstrated improved lesion localization, multiclass classification, and visual interpretability through adaptive augmentation together with SHAP and Grad-CAM explanations; however, these methods continued to rely primarily on benchmark datasets and exhibited limited robustness for minority lesion categories⁴². Several survey articles have summarized these developments, emphasizing both the remarkable progress achieved by modern deep learning techniques and the persistent challenges associated with computational efficiency, explainability, dataset dependency, and clinical validation^43,44,45. Table 1 summarizes some recently published work utilizing deep learning algorithms for Skin lesion classification.

Despite the remarkable advances achieved by recent DL methods in skin lesion classification, most existing approaches remain constrained by high computational complexity, dependence on large annotated datasets, limited model interpretability, and insufficient validation across diverse real-world clinical settings. To overcome these limitations, this paper proposes the EDLF-SLC framework that integrates complementary features extracted from multiple CNN architectures with an SVM classifier for robust and accurate classification. The framework further employs enhanced preprocessing to improve image quality and address class imbalance, while incorporating the LIME technique to provide visual explanations of individual predictions, thereby enhancing model transparency and clinical trustworthiness.

The contributions of this study are threefold. First, the authors propose a robust framework, EDLF-SLC, that integrates advanced convolutional neural networks (CNNs) with a support vector machine (SVM) classifier to achieve accurate and reliable skin lesion classification. Second, the authors implement enhanced preprocessing techniques to address class imbalance and reduce image noise, thereby improving the quality of the input images and the overall performance of the classification model. Third, the authors incorporate the Local Interpretable Model-agnostic Explanations (LIME) method to visualize and interpret individual model predictions by highlighting the image regions that most strongly influence classification decisions, thereby improving the transparency and interpretability of the proposed framework.

Protocol

This study did not involve the recruitment of human participants or the collection of identifiable patient data. All experiments were conducted using the publicly available ISIC 2020 skin lesion dataset obtained from the Kaggle repository; therefore, institutional ethics committee approval and informed consent were not required. All the materials used in this study are in the Table of Materials.

Feature extraction and fusion
Skin lesion images were processed using multiple pretrained CNN models. Deep feature vectors extracted from the selected pretrained CNN architectures are concatenated to construct a unified feature representation for each skin lesion image. The adopted fusion strategy preserves complementary visual information learned by different CNN architectures, allowing the subsequent SVM classifier to exploit both low-level texture characteristics and higher-level semantic representations. Although dimensionality reduction techniques such as principal component analysis or mutual information-based feature selection may reduce feature dimensionality, the complete fused representation was intentionally retained because the fused feature space remains computationally tractable for the employed RBF-SVM classifier while preserving complementary diagnostic information extracted from the heterogeneous CNN backbones.

Classification
The fused feature vectors were used to train an SVM classifier with an RBF kernel. Hyperparameter optimization techniques were employed to determine the optimal classification settings. The classifier then categorized skin lesions into their respective classes.

Explainability
To improve interpretability, LIME was applied to generate visual explanations highlighting the image regions that contribute most significantly to the classification outcomes. These explanations would assist clinicians in understanding and validating the model's decisions.

Evaluation plan
The proposed framework was evaluated using a benchmark skin lesion dataset. Performance was assessed using Accuracy, Precision, Recall, and F1-Score. Comparative experiments were conducted against existing machine learning and deep learning approaches to demonstrate the effectiveness of the proposed framework.

Expected outcomes
The study was expected to produce an accurate and interpretable skin lesion classification system. Preliminary experimental results indicate that EDLF-SLC achieves an accuracy of 88.0%, precision of 89.0%, recall of 87.0%, and F1-score of 88.0%, outperforming several competing methods. The integration of LIME explanations was anticipated to enhance the trustworthiness and facilitate the adoption of AI-assisted diagnostic systems in clinical practice.

Significance
This research contributes to the growing field of explainable healthcare AI by addressing both performance and transparency challenges in skin lesion diagnosis. The proposed framework has the potential to support dermatologists in making more reliable and interpretable diagnostic decisions, ultimately improving patient care. Moreover, in this section, the authors provided information regarding materials and the methodology that has been applied within this research study. Specifically, the authors started with the description of the dataset used for the experiments and moved forward to the detailed discussion of the Explainable Deep Learning Framework for Skin Lesion Classification.

Dataset
The presented work relies on the publicly available ISIC 2020 dataset (International Skin Imaging Collaboration), which can be accessed on the Kaggle website (https://www.kaggle.com/datasets/fanconic/skin-cancer-malignant-vs-benign). The ISIC repository has been recognized as one of the best resources in the field of dermatological imaging, since it provides a large variety of dermoscopic images of different types of skin lesions, as shown in Figure 1. Importantly, the dataset includes images of both benign and malignant conditions, making it useful for performing binary classification tasks for skin cancer ⁴⁶. There are 3,297 images in the current dataset, out of which 1,800 are benign and 1,497 are malignant. For more efficient experiments, the data had to be split into a training set and a testing set. In particular, there are 2,637 training images, including 1,440 benign and 1,197 malignant, while the testing set consists of 660 images, of which 360 are benign and 300 are malignant. A summary of the dataset is shown in Table 2.

The proposed EDLF-SLC model
In this paper, an efficient and understandable solution for classifying skin lesions into benign and malignant categories has been proposed, using a novel explainable deep learning technique based on a hybrid approach combining the advantages of multiple pre-trained convolutional neural network models. Convolutional neural networks have proven highly effective in extracting high-level semantic features from medical images. Using this capability, the proposed explainable deep learning framework for skin lesion classification (EDLF-SLC) exploits multiple pretrained CNN models to achieve superior performance. To ensure the necessary transparency of the medical decision-making process, LIME has been adopted in the proposed approach to generate local human-readable explanations for the output results. The complete framework can be broken down into three main stages, as illustrated in Figure 2, namely, (1) data preprocessing, (2) feature extraction and classification, and (3) interpretability.

Model architecture and integration
As a preliminary step, seven popular deep learning models (MobileNetV2, ResNet50, EfficientNetB0, Xception, InceptionResNetV2, NASNetLarge, and MobileNet) were selected and evaluated on the validation dataset separately, and the four most effective models (ResNet50, EfficientNetB0, MobileNet, and Xception) have been chosen for the feature extraction stage. In the proposed framework, each input image x has been independently passed through the selected pretrained models. The last fully connected layer has been removed from each of these models, and feature vectors have been obtained from the previous layers. fResNet50(x), fEfficientNetB0(x), fMobileNet(x), and fXception(x) refer to the output feature vectors from each of the above-mentioned models. By concatenating those feature vectors, a new aggregated feature vector F(xi) is obtained:

F(x_i) = [f_{(ResNet50)(xi)};f_{(EfficientNetB0)(xi});f_{(MobileNet)(xi)};f_{(Xception)(xi)}] (1)

Afterward, F(x_i) has been passed through an SVM classifier with a radial basis function (RBF) kernel to obtain the classification result. The whole algorithm for the proposed framework is presented in Supplementary File 1.

The proposed framework adopts direct feature-level fusion because each pretrained CNN architecture captures different discriminative characteristics of skin lesions through its distinct network architecture and learned feature hierarchy. Consequently, concatenating these heterogeneous feature representations preserves complementary information that contributes to more comprehensive lesion characterization before classification. Although dimensionality-reduction techniques such as principal component analysis (PCA) or mutual information-based feature selection can reduce the dimensionality of the fused representation, the complete fused feature vector was intentionally retained because its dimensionality remains computationally tractable for the adopted RBF-SVM classifier while preserving complementary diagnostic information extracted by the different CNN backbones. This design, therefore, prioritizes retaining complementary deep representations rather than eliminating potentially informative features before classification.

Data preparation
Data preparation has included preprocessing and splitting the data set into training and testing sets. Preprocessing is aimed at enhancing the quality of data by eliminating inconsistencies, deleting duplicate elements, formatting input images, and performing feature scaling for stable training of a good model. All images have been normalized to the range of [−1, 1] by means of the Z-score normalization:

(2)

where µ and σ stand for the mean value and the standard deviation of the corresponding image, respectively. The data set has been divided into two subsets, i.e., a training set (70.0%) and a testing set (30.0%).

Data augmentation
To avoid overfitting and increase the generalization ability of the model, some augmentations have been used for the training set. Image augmentations involve modifying images in order to artificially expand the data set without changing essential features of medical conditions. The augmentation pipeline has been built using the Keras Sequential API. Such transformations as random horizontal flipping, rotation within a small angle, resizing, scaling, brightness/contrast adjustment, zooming, and some minor movements have been used. Representative examples of augmented images are illustrated in Figure 3.

Pretrained models
Several pretrained deep learning models have been adopted in the proposed framework for extracting features from input images. Those models include MobileNet, ResNet50, EfficientNetB0, Xception, NASNetLarge, Inception-ResNet-v2, and MobileNetV2. MobileNet and MobileNetV2 are lightweight deep learning models for efficient computations using depth-wise separable convolutions. ResNet50 employs the mechanism of residual connections for training the model, thus avoiding the problem of vanishing gradients during training. EfficientNetB0 strikes a balance between efficiency and precision through the approach of compound scaling. Xception extends the Inception network using depthwise separable convolutions. NASNetLarge involves the use of neural architecture search to build optimal deep neural network structures, and Inception-ResNet-v2 builds a hybrid of the Inception model and the residual connections mechanism. Feature vectors extracted from ResNet50 (1,024 features), EfficientNetB0 (1,280 features), MobileNet (1,024 features), and Xception (2,048 features) are concatenated to produce a unified 5,376-dimensional representation. This fusion strategy was selected to preserve complementary representations learned by the different CNN architectures rather than reducing the representation prior to classification. The resulting feature vector is subsequently supplied to the RBF-SVM classifier, which determines the optimal nonlinear decision boundary within the fused feature space.

Explainable AI model (LIME)
In order to provide some local interpretability of the model predictions, the authors adopt the method of generating localized human-readable explanations for each particular model prediction, called LIME⁴⁷. For any input image, LIME first partitions it into smaller units called superpixels. Perturbations are then generated from the original image, and the predictions from the deep neural model are estimated for each of the perturbations. Then, a weighted linear model is fitted on the perturbations using weights depending on their proximity to the original input instance. After fitting the linear model, the feature importance scores are estimated, indicating the role of each superpixel in the prediction of the final label. Those importance scores are visually highlighted in the final explanation image. The LIME algorithm is shown in Supplementary File 2.

Results

Performance assessment of the developed classification framework is based on traditional evaluation measures stemming from the confusion matrix. A confusion matrix allows for obtaining a complete understanding of classifier performance through the representation of the number of right and wrong classifications made for each class defined. In this way, all types of errors can be analyzed. For example, both false positives and false negatives are especially important in the diagnosis of diseases. High values of false positives may show a classifier’s high sensitivity, but at the same time, large quantities of false negatives indicate its low sensitivity and pose potential health risks. The following performance metrics are utilized in this paper. These are accuracy, precision, recall, F1-score, specificity, true positive rate (TPR), and false positive rate (FPR). The formulas of these metrics are listed below:

Accuracy=(TP+TN)/(TP+TN+FP+FN) (3)

Precision =  TP/(TP+FP) (4)

 (5)

(6)

(7)

(8)

In this case, TP shows the number of malignant skin cancers detected by the framework, while TN indicates number of benign lesions. In turn, FP demonstrates a number of wrong decisions when lesions are classified as malignant even though they actually are benign. FN reflects number of wrongly recognized benign samples. With regards to skin cancer classification, minimizing false negatives is extremely important due to potential adverse effects resulting from it.

Performance of baseline models
All computational experiments were conducted in the cloud-based environment of Google Colab. It is worth noting that this platform allows running virtual machine instances using a predefined Ubuntu 20.04 as their operating system. To train baseline models, Python version 3.9 and PyTorch framework version 2.3.1 were used. Additionally, all the computing nodes had identical specifications, namely, an Intel Xeon CPU with a frequency of 2.2 GHz, NVIDIA Tesla T4 GPU, 16 GB RAM, and a storage capacity of 70 GB. Thus, such an experimental setting allowed reducing the variability introduced by different hardware platforms and increasing the reliability and reproducibility of results. A complete summary of the experimental environment can be found in Table 3.

Figure 4 shows learning curves corresponding to 100 epochs of training for the ResNet-50 architecture. Training was performed based on a dataset containing 2,110 images, while the size of the validation dataset was equal to 660 images. As one can see, during training, accuracy constantly increased up to almost 90.0%. At the same time, improvement in accuracy was observed for the first 50 epochs; after this point, the accuracy became almost constant, which can be viewed as a sign of convergence of the model. For validation, the model demonstrated an AUC value equal to 86.0%, thus demonstrating reasonable discriminative properties.

The confusion matrix presented in Figure 5 characterizes the performance of the ResNet-50 model in terms of classification accuracy in the case of a test set with 660 images. During evaluation, the model correctly recognized 310 out of 360 benign samples and 239 out of 300 malignant samples. However, a relatively high number of malignant samples were classified incorrectly, which led to a relatively large value of false negatives. This result should be considered in more detail due to the serious implications of this type of mistake when using the classifier in practice. In total, the model reached an accuracy of 83.0%, with a precision of 83.3%, a recall of 83.3%, and an F1-score of 83.3%.

Table 4 contains a detailed description of performance characteristics for the ResNet-50 model in terms of both classes. The classifier showed relatively balanced behavior in terms of precision, for benign samples, its value was equal to 83.0%, while malignant samples yielded a precision of 84.0%. Similarly, recall values reached 88.0% for benign lesions and 78.0% for malignant ones. Support in the table represents the number of samples corresponding to each class.

Proposed EDLF-SLC model
Figure 6 illustrates the training and validation AUC of the proposed model over 300 epochs. The AUC metric reflects the model’s ability to distinguish between benign and malignant classes, with higher values indicating better discriminative performance. As observed, the training AUC increases steadily throughout the training process, reaching a maximum value of 1.00 at approximately epoch 200. The validation AUC also improves progressively during the initial training stages; however, it begins to plateau after approximately 150 epochs, suggesting convergence of the model. The final validation AUC of 0.95 demonstrates strong generalization capability and effective class separability.

The confusion matrix of the proposed model, presented in Figure 7, shows that the model correctly classified 299 benign samples and 272 malignant samples, while misclassifying 28 benign cases as malignant and 61 malignant cases as benign. In total, the model achieved 571 correct predictions and 89 misclassifications. Notably, the higher number of false negatives (malignant cases misclassified as benign) highlights a critical limitation, as such errors may have significant clinical implications. Nevertheless, the overall classification performance remains robust. Unlike feature-selection approaches that intentionally remove dimensions before classification, the proposed framework preserves the complete fused deep representation generated by multiple heterogeneous CNN architectures. This design was adopted because each backbone extracts complementary lesion characteristics, and retaining the complete representation allows the SVM classifier to exploit the combined discriminative information without excluding potentially informative features. Consequently, the adopted feature fusion strategy prioritizes complementary representation learning while maintaining computational feasibility.

Figure 8 presents representative examples of classification outcomes produced by the proposed model. The results demonstrate that the model assigns high confidence scores to correctly classified instances. Specifically, benign cases exhibit high predicted probabilities (e.g., 99.84% and 99.85%), while malignant cases also show strong confidence levels (e.g., 99.98% and 99.96%). These findings indicate that the model can effectively differentiate between benign and malignant lesions, even in visually challenging scenarios. To enhance interpretability, the LIME framework is employed to generate localized explanations for model predictions, as shown in Figure 9A–D. LIME approximates the complex decision boundary of the model using a locally interpretable linear model. For each input image, the method generates perturbed samples by selectively masking superpixels and evaluating the corresponding predictions. A weighted linear model is then fitted to these samples, where weights are determined based on proximity to the original input using a radial basis function kernel. The resulting coefficients represent the contribution of each superpixel to the final prediction and are defined as:

(9)

where X_j ∈ {0, 1} denotes the presence or absence of the j-th superpixel, B_j represents the corresponding contribution weight, and B₀ is the baseline prediction. Positive coefficients (B_j > 0) indicate regions contributing to the malignant class, whereas negative coefficients (B_j < 0) correspond to benign features. The LIME-based visualizations, shown in Figure 9B, D, highlight the most influential regions contributing to each classification decision. For benign lesions, the model emphasizes regions characterized by uniform pigmentation and well-defined borders. In contrast, for malignant cases, the highlighted regions correspond to clinically significant features such as irregular borders, color heterogeneity, and atypical textures. The intensity of the highlighted regions reflects the magnitude of the corresponding coefficients B_j.

These results demonstrate that the EDLF-SLC model focuses on clinically meaningful features that align with established dermatological criteria, such as the ABCD rule. This alignment enhances the interpretability and trustworthiness of the model, making it more suitable for clinical decision support. Furthermore, Figure 10 presents comparative visualization results obtained using additional explainability techniques, including Grad-CAM, HiResCAM, GradCAM++, XGradCAM, LayerCAM, EigenGradCAM, and EigenCAM, further validating the consistency of the identified salient regions across different methods. With respect to the performance of the proposed EDLF-SLC model and seven baseline models post-training for 100 epochs, a comparison can be seen in Table 5. EDLF-SLC obtained a competitive performance of 0.88 on each metric evaluated compared with the baseline architectures evaluated based on the experimental context. EfficientNetB0 and ResNet50 had good results as well, with accuracies of 0.85 and 0.83, respectively. Conversely, Inception-ResNetV2 had the worst classification performance of the evaluated models. On the other hand, despite relatively lower classification accuracy, its computational efficiency was still comparable to that of the baseline models. The proposed EDLF-SLC model demonstrated a competitive performance relative to the evaluated baseline models under the adopted experimental conditions.

To assess the performance of the proposed framework relative to existing approaches, Table 6 presents a comparison with representative state-of-the-art methods for skin lesion classification. For instance⁴⁷, reported an accuracy of 0.86, while⁴⁸ employed an ensemble-based model that achieved similar accuracy but lower precision, recall, and F1-score. Recent studies based on ensemble learning and multiple CNN architectures^25,49 reported accuracies of up to 0.87. Under the adopted experimental conditions, the proposed EDLF-SLC framework achieved an accuracy of 0.88 while using a unified explainable architecture without requiring additional components such as lesion segmentation models.

DATA AVAILABILITY
The data that support the findings of this study are openly available in Kaggle at https://www.kaggle.com/datasets/fanconic/skin-cancer-malignant-vs-benign.

Figure 1: Representative dermoscopic images from the ISIC dataset illustrating the two diagnostic classes used in this study. Samples are labeled as benign (0) and malignant (1), highlighting the variability in lesion morphology, color, and texture across the dataset. Please click here to view a larger version of this figure.

Figure 2: Complete workflow of the proposed EDLF-SLC framework. The protocol begins with ISIC 2020 image acquisition, followed by preprocessing, data augmentation, dataset partitioning, deep feature extraction using four pretrained CNN architectures, feature fusion, SVM classification, performance evaluation, and LIME-based explanation generation. Please click here to view a larger version of this figure.

Figure 3: Examples of augmented skin lesion images generated using data augmentation techniques. This includes horizontal and vertical flipping, rotation, scaling, and brightness adjustment. These transformations increase dataset diversity, improve model robustness, and reduce the risk of overfitting during training. Please click here to view a larger version of this figure.

Figure 4: Training and validation receiver operating characteristic area under the curve (ROC-AUC) of the ResNet-50 model over 100 training epochs. The blue curve represents the training AUC, whereas the orange curve represents the validation AUC. The curves demonstrate rapid convergence during the initial training epochs, followed by stable optimization with consistently high validation performance, indicating effective learning and satisfactory generalization on the validation dataset. Please click here to view a larger version of this figure.

Figure 5: Confusion matrix of the ResNet-50 model on the test dataset. Rows represent the true class labels (0 = benign, 1 = malignant), whereas columns represent the predicted class labels. The diagonal elements indicate correctly classified samples (310 benign and 239 malignant), while the off-diagonal elements represent misclassified samples, including 50 false positives and 61 false negatives. Please click here to view a larger version of this figure.

Figure 6: Training and validation receiver operating characteristic area under the curve (ROC-AUC) of the proposed EDLF-SLC model over 300 training epochs. The blue curve represents the training AUC, whereas the orange curve represents the validation AUC. The model exhibits rapid convergence during the initial training epochs, followed by stable optimization with consistently high training and validation AUC values throughout the remaining epochs. The sustained validation performance demonstrates good generalization capability and stable learning behavior of the proposed EDLF-SLC framework on the validation dataset. Please click here to view a larger version of this figure.

Figure 7: Confusion matrix of the proposed EDLF-SLC model on the test dataset. Rows represent the true class labels (0 = benign, 1 = malignant), whereas columns represent the predicted class labels. The diagonal elements indicate correctly classified samples (299 benign and 272 malignant), while the off-diagonal elements correspond to misclassified samples, including 61 false positives and 28 false negatives. Please click here to view a larger version of this figure.

Figure 8: Sample predictions generated by the proposed EDLF-SLC model. This figure illustrates classification confidence levels for benign and malignant lesions and demonstrates the model’s discriminative capability. Please click here to view a larger version of this figure.

Figure 9: LIME-based local explanations of the proposed EDLF-SLC model. The figure highlights the most influential superpixel regions contributing to the model's classification decisions. (A) Original dermoscopic image of a benign skin lesion. (B) LIME explanation for the benign lesion, with highlighted superpixels indicating the regions that contributed most to the benign prediction. (C) Original dermoscopic image of a malignant skin lesion. (D) LIME explanation for the malignant lesion, showing the discriminative superpixel regions that predominantly influenced the malignant classification. Yellow boundaries delineate the influential superpixels identified by LIME, while gray regions represent areas with minimal contribution to the prediction. Please click here to view a larger version of this figure.

Figure 10: Comparison of class activation mapping (CAM)-based explainability methods applied to the proposed EDLF-SLC model. The figure compares the localization maps generated by GradCAM, HiResCAM, GradCAM++, XGradCAM, LayerCAM, EigenGradCAM, and EigenCAM for the same dermoscopic image. The first panel shows the original input image, followed by the corresponding raw activation maps and heatmap overlays produced by each explainability method. The visualizations illustrate the differences in spatial localization and highlight the image regions that contribute most strongly to the classification decision of the proposed EDLF-SLC framework. Please click here to view a larger version of this figure.

Ref.	Year	Dataset	Data size	Feature extraction	Method	Accuracy
9	2022	HAM10000 dataset	10015 skin lesion images	Color and shape features	ML algorithms and Convolutional NN models	95.1%
19	2023	PH2 dataset	200 annotated images	Asymmetrical, Streaks, Dots/Globules, Regression areas features	ML models	94.0%
30	2024	HAM10000 dataset	10000 images	Color and shape features	S-MobileNet	97.7%
28	2025	ISIC2020 and HAM10000 datasets	ISIC2020 (12,670 images) and HAM10000 (10,015 images)	Color and shape	Residual network-Long Short-Term Memory (R-LSTM50)	95.7% (ISIC2020); 94.2% (HAM10000)
32	2026	Monkeypox Skin Lesion Dataset (MSLD)	770 images	Context and texture	DenseNet121, Xception, InceptionV3 and CBAM	88% (DenseNet121)
39	2025	HAM10000	38,569 images	Context and texture	MobileNet, ResNet50, InceptionV3, and EfficientNet	93.6% (MobileNet)
40	2025	ISIC 2019	2357 images	Context and Spatial	CNN-transformer-based multi-modal DL	98.71%
41	2026	ISIC 2019	25,000 images	Context and texture	TransXV2S	95.26%
42	2025	HAM10000	10,015 images	Color and shape	ViT with YOLOv8	93%

Table 1: Literature comparison. Summary of some recently published work utilizing deep learning algorithms for Skin lesion classification.

Dataset	Benign	Malignant	Total
Train data	1440	1197	2637
Test data	360	300	660
Total data	1800	1497	3297

Table 2: Dataset distribution. Distribution of benign and malignant samples in the ISIC 2020 dataset, including training and testing splits.

Component	Specification
Operating System	Ubuntu 20.04
Programming Language	Python 3.9
DL Framework	PyTorch 2.3.1
Optimizer	Adam
Learning rate scheduling	1e−3
Number of epochs	100/300
CPU	2 vCPU 2.2 GHz
GPU	Tesla T4 (16 GB) × 1
RAM	16 GB
Trainable Parameters	2,430,353 (9.27 MB)
Non-trainable Parameters	133,434,397 (509.01 MB)

Table 3: System specifications. Detailed specifications of the computational environment, including software frameworks and hardware resources utilized for model training and evaluation.

Class	Precision	Recall	F1-score	Support
Benign class	0.83	0.88	0.85	360
Malignant class	0.84	0.78	0.81	300
Accuracy	-	-	0.832	660
Macro avg	0.84	0.83	0.83	660
Weighted avg	0.84	0.83	0.83	660

Table 4: Classification report. Classification performance of the ResNet-50 model on the test dataset, including precision, recall, F1-score, and support for each class.

Model	Accuracy	Precision	Recall	F1 Score	Time (s)
NASNetLarge	0.77	0.76	0.77	0.76	2002.47
EfficientNetB0	0.85	0.85	0.85	0.85	734.8
Xception	0.8	0.81	0.8	0.8	732.23
ResNetV2	0.63	0.64	0.63	0.611	1072.34
MobileNet	0.79	0.8	0.79	0.79	629.29
MobileNetV2	0.77	0.79	0.77	0.77	660.5
ResNet50	0.83	0.83	0.83	0.83	749.77
EDLF-SLC	0.88	0.89	0.87	0.88	3279.77

Table 5: Model performance comparison. Comparative performance of the proposed EDLF-SLC model and baseline deep learning models across accuracy, precision, recall, F1-score, and computational time.

Model	Accuracy	Precision	Recall	F1-Score
ResNet-18 [25]	0.85	0.85	0.85	0.85
ResNet-50 with SVM [50]	0.87	0.88	0.98	0.93
Hybrid DL models + SVM [48]	0.86	0.84	0.8	0.84
Hybrid DL models + SVM [49]	0.86	0.8	0.6	0.68
Proposed EDLF-SLC	0.88	0.89	0.87	0.88

Table 6: Model comparison metrics. Performance comparison between the proposed EDLF-SLC framework and recent state-of-the-art approaches for skin lesion classification.

Supplementary File 1: Algorithm 1. Workflow of the proposed EDLF-SLC framework, outlining data preprocessing, deep feature extraction using multiple CNN architectures, SVM-based classification, and LIME-based explainability for skin lesion prediction. Please click here to download this file.

Supplementary File 2: Algorithm 2. Workflow of the LIME-based local explanation process, illustrating superpixel generation, perturbation sampling, surrogate model construction, and visualization of the image regions contributing most to the classification decision. Please click here to download this file.

Discussion

The proposed explainable deep learning framework for skin lesion classification (EDLF-SLC) demonstrates that combining complementary deep feature representations with explainable artificial intelligence can improve both classification performance and model transparency. By integrating multiple pretrained CNN architectures (ResNet50, EfficientNetB0, MobileNet, and Xception) for feature extraction and employing a Support Vector Machine (SVM) for final classification, the framework exploits the strengths of different feature extractors to generate richer and more discriminative lesion representations than those obtained from a single backbone network. Furthermore, the integration of local interpretable model-agnostic explanations (LIME) provides localized visual explanations, enabling clinicians to understand the image regions that contribute most to each prediction and increasing confidence in the model's diagnostic decisions.

Experimental results demonstrate that EDLF-SLC achieves competitive performance compared with baseline CNN models, including MobileNet, Xception, and ResNet50, highlighting the effectiveness of complementary feature fusion and SVM-based classification. Comparison with recent state-of-the-art approaches further confirms the competitiveness of the proposed framework. For example, two studies^48,49 reported accuracies of approximately 0.86 using ensemble-based methods, while other hybrid CNN-SVM frameworks^{25,50,51,52,53} achieved accuracies up to 0.87 but relied on more complex architectures and additional preprocessing or segmentation modules. In contrast, the proposed framework achieves an accuracy of 0.88 using a comparatively streamlined architecture without requiring explicit lesion segmentation, while simultaneously providing interpretable predictions through LIME. These results indicate that combining complementary CNN feature extractors before SVM classification can enhance diagnostic performance while maintaining architectural simplicity and transparency.

Although the proposed framework improves classification accuracy and interpretability, this improvement comes at the expense of increased computational cost. The total training time of EDLF-SLC is approximately 3279.77 s, considerably higher than individual CNN models such as ResNet50 (749.77 s) and MobileNet (629.29 s). This additional computational overhead results from training multiple pretrained CNNs and performing feature fusion before classification. Such a trade-off is commonly observed in hybrid and ensemble learning approaches, where improved predictive performance is achieved at the cost of higher computational requirements. Nevertheless, in clinical decision-support systems, diagnostic accuracy, robustness, and reliability are generally considered more important than training efficiency because they directly affect patient outcomes.

Another important advantage of the proposed framework is its explainability. Unlike conventional deep learning models that often function as black boxes, EDLF-SLC incorporates LIME to provide case-specific visual explanations of the prediction process. These explanations help clinicians verify whether the model focuses on clinically relevant lesion regions, thereby improving transparency, supporting clinical interpretation, and facilitating trust in AI-assisted diagnosis. Consequently, the proposed framework offers a practical balance between predictive performance and model interpretability, making it a promising computer-aided decision support tool for skin lesion classification.

Despite these encouraging results, several limitations should be acknowledged. First, the framework was evaluated using only the publicly available ISIC dataset, and its generalization capability across images acquired under different clinical conditions, imaging devices, and patient populations remains to be validated. Second, employing multiple pretrained CNN architectures inevitably increases computational complexity and training time compared with single-backbone models. Third, fixed hyperparameter settings were adopted throughout the experiments, and further optimization may improve performance and adaptability across different datasets. Finally, although LIME enhances model transparency, its explanations are local and perturbation-dependent, meaning that explanation consistency may vary between runs and should be further validated with dermatology experts before routine clinical deployment.

Future research will focus on validating the proposed framework using multiple large-scale and multi-center skin lesion datasets, optimizing computational efficiency through lightweight feature fusion and model compression techniques, investigating advanced hyperparameter optimization strategies, and extending the framework to multiclass skin lesion classification. Moreover, integrating additional explainable AI techniques and conducting prospective clinical evaluations with dermatologists will further strengthen the reliability, interpretability, and practical applicability of the proposed framework in real-world clinical environments.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
ISIC 2020 Skin Lesion Dataset	International Skin Imaging Collaboration (ISIC)	ISIC 2020 Challenge Dataset	Dermoscopic image dataset used for model development and evaluation.
Keras	Keras Team	Integrated with TensorFlow	Construction and training of deep learning models.
LIME (Local Interpretable Model-Agnostic Explanations)	Ribeiro et al.	Python package	Generation of local visual explanations for model predictions.
Matplotlib	Matplotlib Developers	Latest compatible version	Visualization of learning curves, confusion matrices, and evaluation plots.
NumPy	NumPy Developers	Latest compatible version	Numerical computation and array manipulation.
NVIDIA Tesla T4 GPU	NVIDIA Corporation	16 GB VRAM	Model training and inference acceleration.
Pandas	Pandas Development Team	Latest compatible version	Data processing and experimental result management.
Pretrained CNN Models	TensorFlow/Keras Applications	ResNet50, EfficientNetB0, MobileNet, Xception	Deep feature extraction from dermoscopic images.
Python	Python Software Foundation	Version 3.9	Programming language used for implementation.
PyTorch	PyTorch Foundation	Version 2.3.1	Deep learning framework used for feature extraction and model implementation.
Scikit-learn	Scikit-learn Developers	Latest compatible version	Support Vector Machine (SVM), evaluation metrics, and data preprocessing.
Support Vector Machine (RBF Kernel)	Scikit-learn	SVC	Classification of fused deep feature representations.
TensorFlow	Google LLC	Version 2.x	Deep learning framework for model training and inference.
Ubuntu Operating System	Canonical Ltd.	Ubuntu 20.04	Experimental operating environment.