An approach to the dermatological classification of histopathological skin images using a hybridized CNN-DenseNet model

Major highlights

• In this work, we investigated the effects of two pretrained models, DenseNet and ResNet, by hybridizing them with a natural CNN-based architecture on the HAM10000 dataset, which consists of images of different skin diseases.

• We also compared the accuracy of our model after training it on another dataset, PAD-UFES-20, with the accuracy of other published works. The model achieved an accuracy of 95% on the HAM10000 dataset and 91% on the PAD-UFES-20 dataset.

• The proposed model is more accurate than models in other recently published works, such as AlexNet+VGGNet, GoogLeNet+AlexNet, and ResNet50+InceptionV3.

• The model uses histopathological skin images as a dataset and incorporates a hybridized CNN-DenseNet model to automate the diagnosis of skin diseases. Its accuracy can be adapted to further the goals of more explicable diagnostic and clinical applications because its accuracy is equivalent to that of actual clinical professionals working in the real world; the same work can be performed in drastically less time, thereby accelerating the diagnosis.

This research introduces a groundbreaking approach to skin disease diagnosis by innovatively merging a hybrid CNN-DenseNet model. Unlike traditional methods reliant on subjective human assessments, our model represents a pioneering advancement by integrating the strengths of both CNNs and the pretrained DenseNet model. This integration enables the creation of a robust framework capable of intricate feature extraction and comprehensive analysis of dermatoscopic images, surpassing conventional diagnostic accuracy. The unique fusion of these models enhances the learning process, allowing for more nuanced discrimination between various skin diseases, including melanocytic nevi, melanoma, benign keratosis-like lesions, basal cell carcinoma, actinic keratoses, vascular lesions, and dermatofibroma.

Dataset description

The dataset that we chose to construct the hybridized model is known as the HAM10000 dataset, which was created by the ViDIR Group, Department of Dermatology, Medical University of Vienna to detect melanoma using skin lesion analysis in a challenge hosted by the International Skin Imaging Collaboration (ISIC) in 2018 (ViDIR Group, 2018).

The aforementioned dataset comprises dermatoscopic images from various populations that were collected and archived using various modalities (Codella et al., 2018b). There were 10,015 dermatoscopic images in total and 1 ground truth response CSV file that contained the header row and 10,015 corresponding response rows. Each lesion was grouped by picture, and the diagnosis confirmed the type according to 10,015 entries. The lesion identifier for each value is defined under the lesion_id column, while the diagnosis confirmation type values are given under the diagnosis_confirm_type column. Even though the images may have been acquired at various camera angles, under various lighting circumstances, and at various times over the patient’s course of treatment, they all reveal the same primary lesion. Images that follow a stricter technique for diagnostic confirmation are often more challenging for human expert doctors to use for assessment, especially when the images are benign (Alenezi, 2019). The diagnostic confirmation approaches used were “single-image expert consensus”, “serial imaging showing no change”, “confocal microscopy with consensus dermoscopy”, and “histopathology”, in increasing order of rigor, as applied to instances in the current dataset. For instance, an NV (melanocytic nevus) image diagnosed by “histopathology” would typically be a more ambiguous or difficult case for human experts than an “NV” (melanocytic nevus) image diagnosed by “single-image expert consensus” alone, as the former image requires a more invasive procedure for diagnosis (Wang et al., 2023; Yeoh et al., 2023). To obtain the ground truth labels for the dataset, the dermatologists individually reviewed the dermatoscopic images, providing their initial diagnoses based on visual features and patterns. Through thorough discussions and deliberations, a consensus diagnosis was reached by reconciling any differences or disagreements among the experts, ensuring dataset accuracy and reliability. This rigorous process of expert annotation and consensus building adds credibility to ground truth labels, providing valuable insights into the expertise involved in creating the dataset (Yap, Yolland & Tschandl, 2018; Velasco et al., 2023).

As Fig. 2 shows, five samples of each kind of skin disease were obtained for visualization purposes. These images are termed dermoscopic lesion images and are uniquely identified within the dataset with the help of a seven-digit unique identifier. These lesion images were extracted from a range of dermatoscope types, including all anatomic sites (excluding mucosa and nails).

Data visualization and processing

To gain insights into the dataset for model construction, the initial steps involved transforming the CSV dataset into a Pandas DataFrame dataset, followed by an assessment of missing values using the isnull() method. Subsequently, the data types across all fields were analyzed to ensure consistency and accuracy. Visualizations through a Jupyter Notebook facilitated both the data cleaning and exploratory data analysis (EDA) stages.

The dataset included information on seven cell types, with melanocytic nevi exhibiting the highest prevalence, as shown in Fig. 3. Moreover, the technical validation field primarily consisted of histopathological diagnoses. Figure 4 highlights specific body regions, such as the back, lower extremity, and trunk, as areas significantly affected by skin cancer.

The age and sex distributions within the dataset were visualized utilizing Matplotlib (Figs. 5 & 6). Notably, a predominant male predominance was observed, particularly in the age range of 30 to 60 years.

Data splitting, normalization and label encoding

With the EDA performed, we familiarized ourselves with the dataset we were working with. In the next step, we decided to split the data into an 80:20 ratio, with 80% of the data split used for training and the remaining 20% of the data used for testing.

Next, we converted the data into NumPy arrays using the np.asarray() method and calculated the mean and the standard deviation. Then, we normalized the data by subtracting them from their mean values and dividing them by their standard deviation.

Afterward, we encoded the labels to one-hot vectors and split the training set into two parts: 10% of the training set was divided into a validation set on which the model was evaluated at a later stage, and the remaining 90% of the training set was used to train the model.

Pretrained model: DenseNet

DenseNet is a popular pretrained model in deep learning that has demonstrated remarkable performance in various image classification tasks. The architecture of DenseNet is characterized by dense connectivity, where each layer is connected to every other layer in a densely connected feedforward manner. This design facilitates feature reuse and gradient flow, enabling the model to learn effectively and capture intricate patterns and dependencies within images. By leveraging this dense connectivity, DenseNet has shown impressive results in image classification tasks, achieving state-of-the-art performance on benchmark datasets such as ImageNet.

The model’s ability to capture fine-grained details and its efficient use of parameters make it particularly suitable for tasks that require accurate and robust classification, including dermatological classification. The pretrained DenseNet model serves as a valuable starting point for transfer learning, allowing researchers and practitioners to leverage its learned representations and adapt them to specific domains or tasks, thereby reducing the need for extensive training from scratch and accelerating the development of effective models for image classification (IEEE, 2020a; Srinivasu et al., 2021).

Hence, we decided to hybridize their natural CNN with the pretrained DenseNet to leverage its advantages for even more improved accuracy in the classification of skin diseases. The pretrained DenseNet served as a valuable component in this hybrid approach, providing a strong foundation for learning features that can be fine-tuned and further optimized for the specific skin disease classification task (Zhang et al., 2019; Tajbakhsh et al., 2016).

We loaded the DenseNet-121 architecture from the Keras application module with pretrained weights. The include_top parameter is set to false to exclude the fully connected layers of the original DenseNet model. The input shape is defined as (75,100,3) to match the desired dimensions of the input images. The model’s output is obtained from the DenseNet121 base model, followed by a global average pooling layer to reduce spatial dimensions. Two fully connected layers with 1,024 units and a rectified linear unit (ReLU) activation function are added to capture more complex features. Finally, a dense layer with the number of classes in the skin disease classification task and a softmax activation function are added to generate the predictions. The model is compiled with the Adam optimizer, a learning rate of 0.0001, and the categorical cross-entropy loss function. The accuracy was chosen as the metric for assessment.

The model is then trained on 75 epochs using the fit() function with a batch size of 32, specifying the training data, target labels, and validation data for monitoring the model’s performance during training, thus hybridizing the natural CNN with the pretrained DenseNet model and significantly increasing the accuracy of the model by approximately 7%.

Pretrained model: ResNet

ResNet, also known as the residual neural network, is a powerful deep learning architecture renowned for its success in image classification tasks. It introduces the concept of residual learning, which allows for the training of exceptionally deep neural networks by mitigating the vanishing gradient problem. ResNet’s architecture incorporates skip connections that enable information to bypass certain layers, facilitating the flow of gradients during training. These skip connections also aid in capturing fine-grained details and alleviating the degradation problem associated with deeper networks. In the realm of skin disease classification, ResNet has demonstrated impressive performance. Its ability to extract intricate features and learn hierarchical representations from medical images has proven valuable in identifying and distinguishing various skin conditions. By leveraging its deep and residual structure, ResNet has shown remarkable accuracy and robustness, thereby demonstrating its potential as a valuable tool in diagnosing and classifying skin diseases.

Natural convolutional neural networks

We applied four approaches to construct the model for this classification problem. The first approach involved the use of a natural CNNs.

We implemented the sequential model from input to output, passing through a series of neural layers consecutively. For clarity, the model is divided into the following parts:

1. Input shape: The input shape was taken as (75, 100, 3), which corresponded to images with a height of 75 pixels, width of 100 pixels, and three color channels (RGBs).

2. Convolutional layers: The first convolutional layer had 32 filters with a kernel size of (3, 3) and used the ReLU activation function. The padding was set to ‘same’, which meant that the output feature maps would have the same spatial dimensions as the input. The second convolutional layer had 32 filters with a kernel size of (3, 3) and used the ReLU activation function. A max pooling layer with a pooling size of (2, 2) follows the second convolutional layer. Downsampling was performed by taking the maximum value within each 2 × 2 region. A dropout layer with a rate of 0.25 was applied, which randomly set 25% of the inputs to 0 during training to reduce overfitting.

3. Additional convolutional layers: Two additional pairs of convolutional layers were added, each consisting of a convolutional layer with 64 filters and a kernel size of (3, 3), followed by a ReLU activation function and ‘same’ padding. After each pair, a max pooling layer with a pooling size of (2, 2) was applied for downsampling. Dropout layers with rates of 0.40 and 0.5 were added after the second and fourth convolutional layers, respectively.

4. Flattening and dense layers: The flattened layer was used to convert the 2D feature maps from the previous convolutional layers into a 1D vector, preparing the data for the fully connected layers. A dense layer with 128 units and ReLU activation was added. A dropout layer with a rate of 0.5 was applied. Finally, a dense layer with num_classes of 7 units and a softmax activation function was added to output the predicted probabilities for each class.

Next, we employed the Adam optimizer with a learning rate of 0.001 and compiled the model with the loss function set to categorical cross entropy, which is commonly used for multiclass classification problems. A learning rate reduction strategy was implemented to monitor the validation accuracy and reduce the learning rate if no improvement was observed for a certain number of epochs. This technique helps fine-tune the learning rate during training. Data augmentation was applied to prevent overfitting. Various augmentation techniques were employed, such as randomly rotating images within a certain degree range, zooming images, randomly shifting images horizontally, and randomly shifting images vertically. The model was then trained over 50 epochs, and the results are shown in Table 2.

As Table 2 shows, the natural CNN achieved approximately 75.6% accuracy on the validation set, with a loss of approximately 0.64. On the other hand, it could achieve approximately 76.2% accuracy on the test set, with a loss of approximately 0.65. The experiments were performed on an NVIDIA Tesla P100 GPU provided by Kaggle.

Natural convolutional neural networks hybridized with the pretrained DenseNet121 model

We then proceeded to hybridize the model with a DenseNet model for transfer learning. The DenseNet121 model was initialized with pretrained ImageNet weights, and the parameter include_top was set to false to exclude the final fully connected layers of the DenseNet121 model. Custom output layers were added to the model. The output tensor of the DenseNet121 model was stored in variable x. A GlobalAveragePooling2D layer was introduced to reduce the spatial dimensions of the output tensor. Two dense layers with 1,024 units and ReLU activation were applied. Finally, a dense layer with num_classes units and softmax activation was included to enable classification. We compiled the model using the Adam optimizer with a learning rate of 0.0001, and the categorical_crossentropy loss function was employed. The model was then trained for 75 epochs with a batch size of 32. The validation data were subsequently passed to the validation_data parameter to monitor the model’s performance during training. The results are shown in Table 3.

As shown in Table 3, the model achieved an accuracy of approximately 83.0% on the validation set and 83.67% on the test set. The loss values of 0.932697 for the validation set and 1.018943 for the test set indicate the average discrepancy between the predicted and true values. Overall, the model seemed to perform reasonably well, with accuracy values above 80%. The experiments were performed on an NVIDIA Tesla P100 GPU provided by Kaggle.

Natural convolutional neural networks hybridized with a pretrained ResNet model with early stopping

In the next iterations of our attempt, we decided to shift to the NVIDIA GeForce GTX 1660 Ti GPU, which has a significantly higher memory clock speed and core clock speed and a greater CUDA computing capability.

We experimented with the pretrained ResNet model in this iteration. With the same architecture of the natural CNNs, as in the ‘Conclusion’, the results were slightly different: the accuracy achieved on the validation dataset was approximately 76.8%, and the loss value obtained during the evaluation process on the validation dataset was 0.639936, as shown in Fig. 8.

The challenge for researchers was to address the high value of loss evident in Table 2. Therefore, in the next attempt, we decided to hybridize the natural CNN with a ResNet model and evaluate the results. A ResNet model is instantiated and set to exclude the top classification layers. The model’s output tensor is assigned to the variable x, which undergoes a GlobalAveragePooling2D layer to reduce spatial dimensions. Two dense layers with 1,024 units and ReLU activation are added, followed by a final dense layer with num_classes units and softmax activation for classification. The model is then compiled with an Adam optimizer using a learning rate of 0.0001, categorical cross-entropy loss, and accuracy as the evaluation metrics.

The model is then set to be trained for 75 epochs, but to counter the overfitting problem faced in ‘Natural Convolutional neural networks Hybridized with a Pretrained ResNet model with early stopping’ in Table 3, the EarlyStopping callback is used to monitor the validation loss and restore the weights of the best-performing model. The model monitors the validation loss (monitor = ‘val_loss’) during training, waits for 10 epochs without improvement (patience = 10), and restores the best weights of the model when the early stopping condition is met (restore_best_weights = True). The accuracy achieved on the validation dataset was approximately 80.5%, and the corresponding loss was 0.601681. Similarly, for the test dataset with 802 samples, the evaluation process took approximately 1 s per step. The accuracy on the test dataset was approximately 79.2%, with a loss value of 0.604906.

The results are shown in Table 4.

Even though the accuracy was lower than that offered by the hybridized CNN with DenseNet, this model—a natural CNN hybridized with ResNet—offered a lower loss. Hence, this prompted us to find a solution to balance the two problems of each variation in the hybridized model for classifying skin diseases in the HAM10000 dataset.

Modified natural convolutional neural network hybridized with pretrained DenseNet and the ResNet model with LeakyReLU, the RMSProp optimizer, and fewer epochs

In the final attempt, we performed our experiments on the same NVIDIA GTX 1660 ti GPU. In our final attempt, we decided to modify the architecture of the natural CNNs by adding a few more layers and tuning some of the parameters.

The CNN model was given an input shape of (75,100,3) to handle images with a height of 75 pixels, width of 100 pixels, and three color channels (RGBs). The model was created using the sequential class from Keras. The first layer added to the model was a Conv2D layer with 64 filters, a kernel size of (3,3), and a ReLU activation function. The padding was set to ‘same’ to maintain the spatial dimensions of the input. A batch normalization layer was included after each Conv2D layer to normalize the activations. This process was repeated once with another Conv2D layer. A MaxPool2D layer with a pooling size of (2,2) was added to downsample the feature maps and reduce spatial dimensions. A dropout rate of 0.25 was applied to randomly deactivate 25% of the neurons in the previous layer to prevent overfitting. The patterns of two Conv2D layers, followed by batch normalization, MaxPool2D, and dropout were repeated twice more with increased filter sizes of 128 and 256, respectively. The dropout rate was adjusted to 0.4 and 0.5 for each set of Conv2D layers. After the convolutional layers, a flattened layer was used to convert the 3D feature maps into a 1D feature vector. A dense layer with 512 units and ReLU activation was added, followed by batch normalization and dropout at a rate of 0.5 to regularize the network. Finally, the output layer consisted of a dense layer with num_classes units and a softmax activation function to generate class probabilities.

Once again, the Adam optimizer was used since this optimizer is known for its adaptive learning rate and efficient gradient descent optimization. The model was subsequently compiled with the chosen optimizer, and the loss function was set to categorical_crossentropy. Additionally, metrics for accuracy and mean absolute error (MAE) were specified to evaluate the model’s performance during training. To adjust the learning rate during training, a learning rate reduction callback was set. The validation accuracy was monitored, and the learning rate was reduced by a factor of 0.5 if no improvement was observed after 3 epochs. The minimum learning rate was set to 0.00001. Data augmentation was then applied using the ImageDataGenerator class from Keras. Various augmentation techniques, such as rotation, zooming, and horizontal/vertical shifting, were included to prevent overfitting and increase the ability of the model to generalize.

Finally, the model was fitted to the training data using the fit_generator() function, the number of epochs was set to 50, and the batch size was set to 10. The training data were augmented on the fly using the data generator created earlier. Validation data were obtained to evaluate the model’s performance during training. The verbose parameter was set to 1 to display progress updates, and the steps_per_epoch was calculated based on the number of training samples divided by the batch size. A plot of the model accuracy and loss over the epochs is given in Fig. 9.

Next, we hybridized the model with a DenseNet121 model instantiated with weights pretrained on ImageNet. The include_top parameter was set to false to exclude the fully connected layers at the top of the network. A GlobalAveragePooling2D layer was added to perform global average pooling, reducing the spatial dimensions of the feature maps. Two dense layers with 1,024 units and ReLU activation were subsequently added, each followed by leaky ReLU activation, thus modifying our attempts in ‘Conclusion’ so that the model will leak some positive values to 0 if they are close enough to zero. Finally, a dense layer with ‘num_classes’ units and softmax activation was added to generate class probabilities. EarlyStopping callback was once again defined to monitor the validation loss and restore the best weights of the model when no improvement was observed for 10 consecutive epochs. The optimizer was set to Adam with a learning rate of 0.0001, and the loss function was set to categorical_crossentropy. The model was then trained using the fit() function. The training data were provided along with a specified batch size of 32. The validation data were used to evaluate the model’s performance during training.

Next, we trained the model set for 50 epochs, but the early stopping callback stopped the training at the 12th epoch to prevent overfitting, and we further decided to utilize the ResNet-50 model pretrained on ImageNet.

All the other parameters and hyperparameters for the ResNet portion of the final model were kept the same as those for the DenseNet portion described in the previous paragraph, and the model was trained for 25 epochs before it was stopped by the EarlyStopping callback function.

The evaluation metrics are noted in Table 5. It is evident that the final loss of the model was now lower than that of all three attempts made by us in the previous sections. Additionally, the accuracy of the final model was greater than that of the natural CNN and the CNN hybridized with ResNet. Hence, we proceeded to conclude our experiments with the final model.

Performance evaluation metrics

The overall performance of the hybrid model in classifying skin diseases is given below in Table 6. The table uses a short form for each disease, which is as follows: ‘nv’ for melanocytic nevi, ‘mel’ for melanoma, ‘bkl’ for benign keratosis-like lesions, ‘bcc’ for basal cell carcinoma, ‘akiec’ for actinic keratosis, ‘vasc’ for vascular lesions and finally ‘df’ for dermatofibroma.

To quantitatively assess the performance of our proposed skin disease classification model, we employ several widely used evaluation metrics. These metrics provide a comprehensive understanding of the model’s accuracy, precision, recall, and F1 score, which are crucial in evaluating the effectiveness of the classification system.

1. Precision

Precision measures the accuracy of positive predictions. It is calculated as: Precision = True Positives÷(True Positives  +  False Positives)

2. Recall (Sensitivity)

Recall, also known as sensitivity, measures the proportion of actual positives that were correctly identified by the model. It is calculated as Recall = True Positives÷(True Positives + False Negatives)

3. F1-score

The F1-score is the harmonic mean of precision and recall. It provides a balance between precision and recall and is calculated as follows: F1 score = 2 × ((Precision × Recall)÷(Precision + Recall))

These metrics offer valuable insights into the model’s ability to correctly classify different skin diseases, accounting for both false positive and false negative predictions. In our evaluation, we report these metrics alongside accuracy to present a holistic view of the model’s performance. The use of these metrics is vital in assessing the model’s robustness and effectiveness in real-world scenarios.

The proposed method for skin disease classification using hybrid CNN architectures, particularly leveraging DenseNet and ResNet models, has several strengths and weaknesses:

1. Strengths:

   1. High accuracy: The model achieves a notable accuracy rate, as demonstrated by the evaluation metrics, surpassing traditional methods and even competing with or outperforming recent state-of-the-art models.

   2. Robustness: The hybrid architecture enhances the model’s robustness by addressing issues such as vanishing gradients, feature reuse, and effective representation learning, increasing its interpretability at capturing intricate patterns in dermatological images.

   3. Adaptability: Leveraging pretrained models such as DenseNet and ResNet allows for effective transfer learning, which is especially crucial in medical imaging where large labeled datasets might be limited. This adaptability aids in better generalization to unseen data.

   4. Model efficiency: Dense connections in the architecture reduce the number of parameters needed, optimizing the model’s efficiency and making it more suitable for medical image analysis where computational resources might be constrained.

   5. Comparative performance: Comparative studies against other models in recent research have highlighted the superiority of the proposed hybrid models in terms of accuracy, sensitivity, and specificity, particularly when dealing with diverse datasets and different skin types.

2. Weaknesses:

   1. Computational demand: Hybrid models involving DenseNet and ResNet may demand more computational resources during training and inference due to their depth and complexity, potentially limiting deployment on certain platforms.

   2. Interpretability: Deeper models often sacrifice interpretability due to their complexity, which might be a concern in medical applications where understanding the reasoning behind predictions is crucial.

   3. Overfitting potential: Despite attempts to address overfitting using techniques such as data augmentation and early stopping, the model’s complexity might still lead to overfitting on smaller datasets.

   4. Training time: The training process for deep models might be time-consuming, especially when working with larger datasets and complex architectures, affecting the scalability of the model.

   5. Limited generalization: While achieving high accuracy, the model’s performance might vary according to demographic group or clinical setting, necessitating further validation across diverse populations.

The confusion matrix provided in Fig. 10 shows the performance of a multiclass classification model on a dataset of six classes. The rows of the matrix represent the true labels of the data, and the columns represent the predicted labels. The diagonal entries of the matrix represent the number of correctly classified samples, and the off-diagonal entries represent the number of incorrectly classified samples. The overall accuracy of the model is 80%, calculated by dividing the number of correctly classified samples (300 + 400 + 523 + 38 + 31 + 11) by the total number of samples (600).

The classwise accuracy of the model is as follows:

• Class 0: 60%

• Class 1: 80%

• Class 2: 75%

• Class 3: 100%

• Class 4: 73%

• Class 5: 62%

Overall, the confusion matrix shows that the model performs well, with an overall accuracy of 80%. The model performs exceptionally well on class 1, with 80% of the samples being correctly classified. This finding suggested that the model learned to distinguish class 1 from other classes with a high degree of accuracy. Class 2: The model performs impressively on class 2, with 75% of the samples being correctly classified. The model performs perfectly on class 3, with all of the samples being correctly classified. This finding suggested that the model learned to distinguish class 3 from other classes with perfect accuracy. The model performs well on class 4, with 73% of the samples being correctly classified. Overall, the confusion matrix shows that the model performs well.

Modified CNN hybridized with a pretrained DenseNet and ResNet model on the PAD-UFES-20 dataset

In this subsection, we attempt to train our model on another database and perform another comparison to better understand the performance of the model. We used the PAD-UFES-20 dataset. This dataset was collected along with the Dermatological and Surgical Assistance Program (PAD) of the Federal University of Espírito Santo. This dataset contains 2,298 samples of six different types of skin lesions, namely, basal cell carcinoma (BCC), squamous cell carcinoma (SCC), actinic keratosis (ACK), seborrheic keratosis (SEK), Bowen’s disease (BOD), melanoma (MEL), and nevus (NEV) lesions. Since the model was trained to predict only melanocytic nevi, melanoma, benign keratosis-like lesions, BCC, actinic keratoses, vascular lesions, and dermatofibroma, we first cleaned the dataset and removed samples from the SCC, SEK, and BOD cohorts. The dataset was subsequently checked for null values, and images were resized to 100x75 so that they could be handled by TensorFlow.

Next, we converted the data into NumPy arrays using the np.asarray() method and calculated the mean and the standard deviation. Then, we normalized the data. Afterward, we encoded the labels to one-hot vectors and split the training set into two parts: 10% of the training set was divided into a validation set on which the model was evaluated at a later stage, and the remaining 90% of the training set was used to train the model.

The model in ‘Modified natural convolutional neural network hybridized with pretrained DenseNet and the ResNet model with LeakyReLU, the RMSProp optimizer, and fewer epochs’ achieved an accuracy of 0.910784 with a loss of 0.089216 on the PAD-UFES-20 dataset. Table 7 below shows a comparison of the models’ performances on 2 different datasets, namely, HAM10000 and PAD-UFES-20.

Comparison of model with other published works

In this subsection, we compare our model’s accuracy with those of works published on a topic similar to and subject of this article.

As Table 8 shows, we compare our proposed model with other recently published models in increasing order of accuracy. The table shows the competitiveness of the proposed model when compared to other models proposed by other researchers in recent years. The proposed hybrid CNN-DenseNet model provided an accuracy of 95.7% compared to Harangi (2018), which had an accuracy of 79% and 80%, (Shahin, Kamal & Elattar, 2018), which had an accuracy of 89%, (Hameed, Shabut & Hossain, 2018), which had an accuracy of 91%, and (Maron et al., 2020; Chen et al., 2023), which had accuracies of 95% and 94%, respectively. As a result, this model may be efficiently utilized to automate the detection of illnesses such as melanocytic nevi, melanoma, benign keratosis-like lesions, BCC, actinic keratoses, vascular lesions, and dermatofibroma. Because its accuracy is comparable to that of actual clinical professionals working in the real world, it may also be tailored to serve the aims of further explainable diagnostic and clinical applications (Zhu et al., 2021).