DEMNet-CERNet: A Dual-Ensemble Framework for Automated Skin Lesion Classification Using Multiple Deep Learning Architectures

Melanoma is among the most lethal forms of skin cancer due to its rapid progression and ability to metastasize to other organs. Despite being less common than other types of skin cancers, late diagnosis could result in ineffective treatment and decreased chances of survival. Therefore, early and accurate diagnosis of melanoma is essential in order to facilitate the treatment process.

It is rather difficult to diagnose melanoma due to the similarity in appearance between benign and malignant skin lesions. Dermatologists usually conduct visual and dermoscopic examinations and apply the ABCDE criteria (asymmetry, border irregularity, color variability, diameter, and evolution) when evaluating suspicious lesions. However, the variability in the size, shape, color, structure, and pattern of skin lesions makes the diagnostic procedure difficult. Moreover, the presence of artifacts such as hair occlusion, noise, and uneven illumination can interfere with the identification of specific lesion features in dermoscopic images.

Large databases of dermoscopic images facilitate the application of DL techniques for skin lesion classification. HAM10000 database comprises 10,015 dermoscopic images divided into seven classes representing pigmented skin lesions and is widely used for evaluating the performance of automated classifiers [1]. Current research has demonstrated the strong capabilities of CNNs for the analysis of skin lesions. Esteva et al. [2] implemented an Inception-v3 CNN for skin lesion classification and showed performance similar to dermatologists, while Brinker et al. [3] examined the CNN classification of melanoma and benign nevi. Transfer learning using pre-trained DL models was also applied for improvement of automated melanoma detection as suggested by Kassani and Kassani [4].

Nevertheless, intra-class similarity and inter-class differences in skin lesion characteristics can complicate classification, even for individual DL models. Hence, ensemble-based methods have increasingly been adopted to take advantage of multiple models to enhance the accuracy of classification. Sharma et al. [5] introduced an ensemble that integrates both deep and manually crafted features to classify skin lesions. Kausar et al. [6] considered the problem of multi-class skin cancer classification using modified transfer learning models, whereas Qureshi and Roos [7] proposed an ensemble framework for highly unbalanced dermoscopic data. An ensemble of multi-resolution EfficientNet models along with patient metadata was designed by Gessert et al. [8], and multi-scale image representations using ensemble learning were studied by Tsai et al. [9]. Even though DL, transfer learning, and ensemble frameworks are proven to be effective in skin lesion classification, the selection of complementary models and combining their outputs remains challenging.

This study focuses on the development of a DL-based model for classifying skin lesions through multiple pre-trained architectures and ensemble learning techniques. The method involves image pre-processing and augmentation, which enhance image quality and minimize noise and artifacts. The processed images are then fed into TL models, where pre-trained models are fine-tuned to extract discriminative features. In order to enhance the performance of classification, predictions from multiple models are combined using an ensemble learning technique. Specifically, a performance-based weighting technique has been employed to assign greater weight to models that performed well during validation.

The proposed system consists of two ensemble architectures—Deep Ensemble Model Network (DEMNet) and Convolutional Enhanced Representation Network (CERNet)—designed to capture complementary feature representations. DEMNet combines DenseNet-121, EfficientNet-B0, and MobileNetV2 to achieve dense connectivity, computational efficiency, and lightweight feature extraction. Conversely, CERNet combines ConvNeXt-Tiny, EfficientNetV2-B0, and ResNet50 to create a global contextual understanding of the entire image for both global context and hierarchical feature interpretation.

The selected models are evaluated using standard performance metrics such as accuracy, precision, recall, and F1-score to assess their performance. The present study employs Gradient-weighted Class Activation Mapping (Grad-CAM) to improve model interpretability, enhance transparency, and support clinical decision-making. This method provides a visual explanation for a model's prediction by showing which regions of interest in the input image contributed most to the model's prediction and, thereby enhancing clinical interpretability.

DL techniques, such as CNNs, TL, and ensemble DL models, have been widely employed in existing studies on skin lesion classification. Nevertheless, these approaches often exhibit limited feature representation, reduced generalizability, and lower robustness because of their reliance on individual models.

To address the issues identified by previous studies, this work proposes a novel dual-ensemble DL framework for enhancing skin lesion classification. The key contributions of this study include:

1. Introduction of a novel aggregation framework called DEMNet-CERNet, specifically designed for the automated classification of skin lesions from dermoscopic images.

2. Six contemporary deep CNN architectures are integrated within the framework, with each network contributing complementary features to improve skin lesion classification accuracy.

3. An adaptive weighting approach based on validation AUC is developed to optimally combine predictions from multiple ensemble models.

4. Grad-CAM is employed within the proposed framework to demonstrate how the neural network interprets each classified image by visually highlighting the image regions that contribute most to lesion classification, thereby providing evidence of model interpretability.

The remainder of this paper is organized as follows. Section 2 reviews related work on CNN-based skin lesion classification, transfer learning approaches, and ensemble DL methods. Section 3 defines the proposed methodology, including image preprocessing, DL architectures, and the proposed dual-ensemble strategy. Sections 4 and 5 present the experimental setup and performance metrics. Section 6 presents the results and discussion, including comparative evaluation and Grad-CAM analysis. Finally, Section 7 concludes the paper and discusses future research directions.

3.1 Overview of proposed architecture

The present study introduces a novel Dual Ensemble Melanoma Classification Framework, DEMNet-CERNet, for accurate and robust skin lesion classification using dermoscopic images. The suggested framework assimilates multiple deep CNNs organized into two complementary model groups and combines their predictions through an ensemble strategy. The main objective is to enhance classifier performance by utilizing complementary feature representations generated by different DL models. The framework also enables visualization of the classification decisions made by each model using explainable AI techniques.

Figure 1 illustrates the architecture of the proposed DEMNet-CERNet framework. It consists of five major stages: input data acquisition, data preprocessing, dual-model feature extraction, ensemble classification, and explainable prediction visualization. These stages collectively enable the system to perform reliable melanoma detection across heterogeneous dermoscopic datasets.

1.png

Figure 1. Proposed Deep Ensemble Model Network (DEMNet)-Convolutional Enhanced Representation Network (CERNet) framework

3.2 Dataset description

The development of DL models for skin lesion classification has been significantly facilitated by the availability of publicly accessible dermoscopic image datasets. Among these datasets, HAM10000 is one of the largest and most widely used resources for skin lesion classification. It contains more than 10,000 dermoscopic images from various clinical facilities and includes images from each pigmented skin lesion class. The large size of the dataset makes it valuable for training DL models. Similarly, the ISIC 2016 and ISIC 2017 challenge datasets have played a significant role in advancing automated melanoma detection research. The ISIC challenge datasets come from the ISIC international challenge; both datasets contain dermoscopic images that have been accurately annotated for melanoma classification and lesion segmentation. In addition to providing annotated images, the ISIC challenge establishes standardized benchmarks for evaluating algorithm performance, enabling fair comparisons among different algorithms and supporting the development of robust automated skin lesion classification systems. It is essential that models be trained on multiple different image datasets prior to evaluation. This is due to variations in image acquisition conditions, illumination, and lesion-class distributions that will be encountered in a variety of clinical environments within dermatology.

The proposed dual-ensemble-based methodology was evaluated on three benchmark publicly available dermoscopic image datasets comprising diverse skin lesion types, expert-annotated labels, and widely recognized standards for the development and evaluation of automated DL-based skin lesion classifiers. These datasets are HAM10000, ISIC 2017, and ISIC 2016. Table 2 summarizes the datasets used for melanoma classification, along with sources, number of images, and their classification tasks.

3.2.1 HAM10000 dataset

HAM10000 is a widely used benchmark dataset for skin lesion classification. The HAM10000 dataset contains 10,015 dermoscopic images collected from various clinical sources, including the Department of Dermatology of the Medical University of Vienna, the Skin Cancer Practice in the state of Australia, etc. The database comprises 7 different types of pigmented skin lesions.

For this study, the HAM10000 dataset was converted from a multiclass dataset into a binary classification dataset. Images labeled as melanoma (mel) were assigned to the positive class, whereas all other lesions were assigned to the non-melanoma class. This conversion enables the development of DL models specifically designed for melanoma detection. Because of this, the HAM10000 dataset's size and diversity will continue to serve as a benchmark for the evaluation of automated skin lesion classification systems.

3.2.2 ISIC 2016 dataset

The ISIC 2016 dataset, part of the ISIC Skin Lesion Analysis Challenge, was created to provide a collection of dermoscopy images to facilitate the development of automated melanoma detection algorithms using both clinical dermoscopy images and clinical expert annotations. The primary focus of this dataset is the binary classification of melanoma and non-melanoma lesions based on images collected from multiple clinics around the world. The ISIC 2016 dataset has been an important resource in DL research for the evaluation of melanoma detection models.

3.2.3 ISIC 2017 dataset

The ISIC 2017 dataset is an enhanced version of the ISIC Challenge Dataset and includes lesion segmentation, dermoscopic feature extraction, and disease diagnosis. The dataset comprises three lesion categories: melanoma, seborrheic keratosis, and benign nevi. The ISIC 2017 dataset contains high-quality dermoscopic images with corresponding ground-truth annotations. This enables the development and evaluation of advanced DL methods for both image segmentation and classification tasks. Numerous studies have used this dataset to improve automated melanoma identification systems because the ISIC 2017 dataset contains both annotated lesion masks and diagnostic classifications.

The dataset was converted to a binary classification problem in line with the research objective of detecting melanoma lesions. Melanomas were classified under the positive class, while seborrheic keratosis and benign nevi were categorized under the negative class (non-melanomas). This was chosen in order to allow the model to learn how to classify melanomas against non-melanomas that have a visually similar appearance.

Table 2. Description of datasets

3.3 Data pre-processing

Image preprocessing is an essential step because raw medical images may contain variations in scale and orientation, as well as background interference. Pre-processing dermoscopic images allows for better feature extraction and assists in building improved DL models. Many studies have demonstrated that preprocessing can improve the performance of DL models for melanoma detection. Therefore, before training the DEMNet-CERNet dual-ensemble framework, various preprocessing techniques were applied. One of the limitations of the current work is that the pre-processing does not include dedicated hair removal or image inpainting techniques, and these techniques will be considered for future work to improve the classification performance for images with significant visual artifacts.

3.3.1 Image resizing

The spatial resolution of images varies across the benchmark datasets. To provide a uniform input size for the DL architectures, all images are resized to 224 × 224 pixels, which is a commonly used input size for architectures such as DenseNet, EfficientNet, and MobileNet.

3.3.2 Data augmentation

The scarcity of annotated dermoscopy images presents a major problem while training the DL models. In order to augment the diversity of the dataset and to avoid overfitting, data augmentation methods have been used during the training of the algorithm. Let the original image be denoted as $I$. The augmented image $I^{\prime}$ is generated using transformation function $T: I^{\prime}=T(I)$. A variety of image augmentation operations were incorporated during training to enhance data diversity.

Random Flip. Horizontal and vertical flipping simulate variations in lesion orientation and improve model invariance.

Random Rotation. Dermoscopic lesions may appear at arbitrary angles. Therefore, random rotation was applied to improve rotational invariance.

$\left[\begin{array}{l}x^{\prime} \\ y^{\prime}\end{array}\right]=\left[\begin{array}{cc}\cos \theta & -\sin \theta \\ \sin \theta & \cos \theta\end{array}\right]\left[\begin{array}{l}x \\ y\end{array}\right]$                      (1)

where, $\theta$ represents the rotation angle. Rotation-based augmentation has been shown to significantly improve melanoma classification accuracy in CNN models.

Random Zoom. Random zoom simulates variations in lesion scale and size.

$I^{\prime}(x, y)=I(s x, s y)$                    (2)

where, $s$ denotes the scaling factor that controls the degree of zoom with

$s \in[0.9,1.1]$

Values below 1 represent a slight zoom-out, whereas values above 1 represent a zoom-in. This transformation improves the ability of the DL models to recognize lesions across variations in scale.

3.3.3 Image normalization

Before being fed into the neural network models, the pixel values of the images were normalized to facilitate faster convergence and more stable training. For an image I(x, y), pixel normalisation is expressed mathematically as:

$I_n(x, y)=\frac{I(x, y)}{255}$                 (3)

This operation takes pixel values from the range [0, 255] and converts them into their equivalent normalised pixel value [0, 1]. Pixel normalization is commonly employed in DL pipelines to standardize input data, thereby improving training stability, facilitating faster convergence, and enhancing optimization performance.

3.3.4 Dataset splitting

To develop the models and evaluate their performance objectively, the dermoscopic images were randomly divided into three non-overlapping sets: training, validation, and test sets. The distribution was chosen as 80:10:10 for the training, validation, and test sets, respectively. In particular, this splitting strategy helps prevent data leakage, facilitates model tuning during training, and enables evaluation of the generalization performance of the developed models.

3.3.5 Class balancing

Typically, there is a class imbalance in skin lesions, with the benign class having a very large number of examples compared to the malignant class. Such class imbalance can bias the model toward predicting the majority class. Class weighting was used to eliminate this effect during training. Let $N=$ total number of samples, $N_i=$ number of samples in class $i, C$ is the number of classes.

The class weight is computed as

$W_i=\frac{N}{C \times N_i}$                 (4)

Applying class weights increases the penalty for misclassifying minority-class samples and is intended to improve the classification of melanoma lesions.

3.4 Dual model feature extraction

The major component of the proposed framework is a dual-ensemble architecture comprising two complementary modules, DEMNet and CERNet, each consisting of multiple CNN architectures that extract diverse and discriminative feature representations from dermoscopic images. The two models are referred to as DEMNet and CERNet. Each of these two models contains multiple different CNNs that capture different levels of melanoma characteristics to produce features for skin lesions based on spatial hierarchies and semantics. There have been substantial studies demonstrating that DL-based feature extraction can automatically learn discriminative representations from dermoscopic images. However, there are limitations in using a single training architecture, particularly the poor ability to generalize across heterogeneous datasets. To avoid these limitations, the proposed approach utilizes multiple CNN architectures to learn complementary features of a lesion, thus being able to create more robust classification.

3.4.1 Deep Ensemble Model Network module

The DEMNet focuses on extracting rich hierarchical image features using well-established CNN architectures. TL is applied to leverage pretrained models trained on large-scale datasets such as ImageNet, which improves convergence and feature representation for medical imaging tasks.

DEMNet integrates the following CNN architectures: DenseNet-121, EfficientNet‑B0, and MobileNetV2. Each architecture contributes unique structural characteristics that improve feature diversity.

DenseNet-121 has very dense connectivity between layers, which allows for the reuse of features above and below through efficient gradient propagation. One of the reasons for DenseNet’s effectiveness in medical image classification is its ability to capture very fine-grained spatial information.

EfficientNet-B0 employs a compound scaling approach that automatically adjusts the network’s width, depth, and image resolution, leading to improved performance using fewer parameters. EfficientNet models have demonstrated strong performance in skin lesion classification.

MobileNetV2 is a lightweight architecture designed using depth-wise separable convolutions and inverted residual blocks. Despite its low computational complexity, MobileNetV2 provides effective feature learning capabilities suitable for medical image classification.

This framework makes use of TL through fine-tuning of pre-trained CNNs for the classification of dermoscopic images.

Before model training, image preprocessing and lesion enhancement are performed to improve image quality. Subsequently, image augmentation techniques, including rotation, flipping, and scaling, are applied to increase data diversity and reduce overfitting. The processed images are then fed into the CNN architectures to extract deep feature representations for melanoma classification.

3.4.2 Convolutional Enhanced Representation Network module

CERNet is designed to further improve feature diversity by incorporating modern convolutional architectures capable of capturing complex visual patterns in dermoscopic images. CERNet integrates the following DL models: ConvNeXt‑Tiny, EfficientNetV2‑B0, and ResNet50.

ConvNeXt-Tiny employs modern convolutional design patterns influenced by transformer architectures to enhance how features are represented and to allow the model to scale.

EfficientNetV2-B0 improves training efficiency through progressive learning and optimized scaling, thereby enabling faster training and strong performance in medical image classification.

ResNet50 has been designed with residual connectivity between layers to overcome the vanishing gradient issue that occurs with DNNs. In addition, residual learning facilitates the learning of complex visual features and improves training stability, making ResNet50 suitable for skin lesion classification.

The outputs provided by the DEMNet and CERNet modules contain prediction probabilities for skin lesion classes; these probabilities are ultimately combined using ensemble learning techniques.

The models for DEMNet and CERNet are selected to provide diverse and complementary CNN architectures. DenseNet-121 and ResNet50 capture hierarchical features through dense and residual connections; EfficientNet-B0 and EfficientNetV2-B0 provide efficient feature extraction through compound scaling and optimized training; MobileNetV2 offers computational efficiency through depthwise separable convolutions; and ConvNeXt-Tiny introduces modern CNN design principles inspired by transformer architectures. This diversity enables the proposed framework to utilize complementary feature representations and enhance classification robustness. Since this work focuses on CNN-based frameworks, transformer-based models are not considered and are left for future investigation.

3.5 Ensemble classification strategy

The dual-model system uses ensemble learning techniques to combine the predictions made by each of the models in the dual system. Ensemble learning can improve classification accuracy by aggregating the outputs of multiple models, thereby reducing prediction variance and enhancing the robustness of the resulting model.

The following ensemble techniques are implemented: Soft Voting, Hard Voting, and Weighted Averaging.

3.5.1 Soft voting

Soft voting combines the predicted class probabilities from all models and assigns the class with the highest average probability as the final prediction. If $P_i(x)$ is the probability that model i will predict the label for input image x, the result of soft voting for the ensemble of N models is:

$P_{\text {soft}}(x)=\frac{1}{N} \sum_{i=1}^N P_i(x)$                   (5)

The final predicted class is obtained as:

$y_{\text {pred}}=\arg \max \left(P_{\text {soft}}(x)\right)$                (6)

Soft voting allows the ensemble to leverage the confidence levels of individual models, improving prediction stability.

3.5.2 Hard Voting

Hard Voting selects the final predicted class label by choosing the majority predicted class label from each of the models in the ensemble. If $y_i(x)$ is the predicted class label returned by model i for input image x, the final predicted class label by the ensemble of N models is:

$y_{\text {pred}}=\operatorname{mode}\left(y_1(x), y_2(x), \ldots, y_N(x)\right)$

This strategy is effective when multiple models produce consistent predictions, reducing the impact of occasional misclassifications.

3.5.3 Weighted average ensemble

The weighted averaging strategy aggregates the outputs of the constituent models by assigning performance-based weights derived from the validation set. Let $w_i$ represent the weight assigned to model $i$, where

$\sum_{i=1}^N w_i=1$               (7)

The weighted ensemble prediction is calculated as:

$P_{ {weighted}}(x)=\sum_{i=1}^N w_i P_i(x)$               (8)

The final prediction is obtained as:

$y_{ {pred}}=\arg \max \left(P_{ {weighted }}(x)\right)$                 (9)

Unlike traditional ensemble models, in which all models contribute equally to the ensemble decision, the proposed framework uses a PAF strategy to combine the predictions of multiple models. Each model was assigned a weight based on its validation AUC, so that models with better validation performance contributed more to the final prediction. The weight for the i^th model was computed as:

$w_i=\frac{A U C_i}{\sum_{i=1}^N A U C_i}$                (10)

where, AUC_i is the validation AUC of the i^th model, N is the number of models in the ensemble, and the sum of weights should be equal to 1. The final prediction was obtained by a weighted average of the class probability outputs.

This PAF mechanism improves classification robustness by prioritizing models demonstrating stronger validation performance.

3.6 Explainable Artificial Intelligence

In order to improve the model interpretability, the proposed framework applies a technique known as Grad-CAM. This technique generates visual heat maps that highlight regions of the image that contribute most strongly to the model's prediction.

Given the feature maps $A^k$ from the final convolutional layer and the gradient of the target class score $y^c$, Grad-CAM computes importance weights as:

$\alpha_k^c=\frac{1}{Z} \sum_i \sum_j \frac{\partial y^c}{\partial A_{i j}^k}$                 (11)

The final Grad-CAM heatmap is generated as:

$L_{\text {Grad-CAM }}^c=\operatorname{ReLU}\left(\sum_k \alpha_k^c A^k\right)$                (12)

Grad CAM  provides a visual representation of the lesion-relevant areas in the dermatoscopic image, which allows clinicians to review how the model reached its final label. Additionally, misclassified examples have been incorporated into the qualitative analysis to provide a more balanced interpretation of the model. These examples demonstrate that incorrect predictions may occur when lesions exhibit visual similarities across classes or have ambiguous boundaries and low contrast, as well as in the presence of artifacts such as hair and variations in image acquisition.

Several performance metrics, including accuracy, precision, recall, F1-score, and specificity, were used to evaluate the developed framework. Accuracy refers to the overall correctness of predictions, while Precision and recall reflect the model's ability to correctly identify melanoma lesions. F1-score represents the balance between precision and recall, while specificity is concerned with the capacity of the model to recognize benign lesions. Therefore, the models’ discriminative abilities were evaluated at different decision thresholds through the use of an ROC curve and an AUC score. These metrics provide a comprehensive evaluation of the performance of DEMNet, CERNet, and the proposed dual-ensemble framework for skin lesion classification.

$Accuracy$ $=\frac{T P+T N}{T P+T N+F P+F N}$                 (13)

$Precision$ $=\frac{T P}{T P+F P}$              (14)

$Recall$ $=\frac{T P}{T P+F N}$                 (15)

$F 1-$ $Score$ $=2 \times \frac{ {Precision} { × } {Recall}}{{Precision}+ {Recall}}$                (16)