Classifying Melanoma in ISIC Dermoscopic Images Using Efficient Convolutional Neural Networks and Deep Transfer Learning

1.png

Figure 1. Skin cancer types

3.1 Convolutional Neural Network (CNN)

The CNN stands out as an innovative paradigm within the realm of multilayer perceptrons (MLP), meticulously designed for the processing of two-dimensional data. CNNs have garnered substantial attention owing to their widespread application and inherent capability to accommodate high-depth networks, making them particularly adept at analyzing image data [26]. The architectural framework of CNNs shares similarities with general neural networks, featuring neurons equipped with weights, biases, and activation functions. A typical CNN structure encompasses a convolution layer employing Rectified Linear Unit (ReLU) activation, succeeded by a pooling layer that serves as the feature extraction stage, culminating in a fully connected layer employing softmax activation for classification purposes.

3.1.1 Convolution Layer

At the core of CNNs lies the fundamental operation of convolution, primarily executed within the convolution layer. Operating as the initial processing layer for input images, this layer utilizes a convolutional filter, typically with dimensions such as 3 × 3, 5 × 5, or 7 × 7, to convolve the image. This convolution process generates multiple feature maps, often referred to as the feature map set [27], as depicted in Eq. (1).

$F_j^{l+1}=\sum_i w_{i j}^l F_i^l+b_j^l$            (1)

Here:

• F represents the j-th feature map of the l-th layer.
• w signifies the convolutional kernel filter.
• b represents the bias.

In the training phase of CNNs employing the Backpropagation technique, convolutional kernel filters (w), responsible for extracting features from input images, undergo initialization with small random values before the commencement of training. Throughout the forward pass of the backpropagation process, these filters are applied to input dermoscopic images, resulting in the generation of feature maps. Each feature map captures unique image patterns. The optimization technique of gradient descent is then deployed to adjust the filter weights, with gradients supplying crucial information on how the filters should be updated to minimize the loss function. The learning rate, recognized as a hyperparameter, dictates the step size during this process. Filters undergo iterative updates across epochs by subtracting a fraction of the gradient from their current values. This iterative process enables the neural network to progressively learn and refine features tailored specifically for the task at hand, which, in this study, is melanoma classification.

Similarly, biases (b) also commence with small random values and are incorporated into convolutional outputs and layer activations during the forward pass. Loss gradients associated with biases are computed, and gradient descent is utilized to update biases, aiming to minimize the loss.

3.1.2 Batch normalization layer

The primary goal of batch normalization is to alleviate issues related to data saturation, enabling neural networks to achieve faster convergence rates and greater resilience against problems associated with parameter initialization [27]. By integrating batch normalization layers between convolutional layers and nonlinearities, such as ReLU layers, the network training process is accelerated, and overall stability is enhanced.

3.1.3 Rectified Linear Units (ReLUs)

Rectified Linear Units (ReLUs) function as activation layers within CNNs, expediting the neural network training phase while minimizing errors. In instances where a pixel's value (x) within an image falls below zero, the ReLU activation promptly sets all pixel values to zero [28], as illustrated in Eq. (2).

$f(x)=\left\{\begin{array}{l}x \text { if } x>0 \\ 0 \text { if } x \leq 0\end{array}\right.$            (2)

The pivotal characteristic of Rectified Linear Units (ReLUs) lies in their ability to introduce non-linearity into deep neural networks, a crucial aspect enabling the network to recognize and comprehend complex patterns and relationships within the input data. This non-linearity is vital for CNNs, as it enables the modeling and learning of intricate features within input images. Additionally, ReLUs play a role in mitigating the vanishing gradient problem, a significant issue often encountered in deep neural networks [28].

3.1.4 Polling layer

The pooling layer provides several advantages, including controlling the size of the output volume on the feature map to prevent overfitting. Commonly positioned after multiple convolution layers within the CNN architecture, this layer performs data reduction through mean- or max-pooling. Mean-pooling calculates the average value, while max-pooling selects the highest value among elements within a small neighborhood.

3.1.5 Fully Connected (FC) layer

In the + data through a flattening process [29]. Following this process, the logistic regression technique can be utilized with softmax activation for the classification of more than two groups.

3.1.6 Optimizer

Hyperparameters play a crucial role in model performance as they require adjustment during model training. In this study, we utilize the Stochastic Gradient Descent (SGD) optimizer to fine-tune the CNN networks. This optimization technique leverages each training sample as a parameter to improve the model's performance by utilizing advanced mathematical functions such as derivatives or subderivatives [30]. The SGD update equation can expressed as follows:

$\theta_{(\mathrm{i}+1)}=\theta_{\mathrm{i}-\alpha} \nabla \mathrm{L}\left(\theta_{\mathrm{i}}\right)+\gamma\left(\theta_{\mathrm{i}}-\theta_{(\mathrm{i}-1)}\right)$              (3)

Here:

• $\theta_{(\mathrm{i}+1)}$ represents the updated parameter vector.
• $\theta_{\mathrm{i}}$ denotes the current parameter vector.
• α'signifies the learning rate.
• $\nabla \mathrm{L}\left(\theta_{\mathrm{i}}\right)$ is the gradient of the loss function with respect to the current parameters $\theta_{\mathrm{i}}$.
• γ controls the momentum and influences how the previous gradient step contributes to the current iteration.
• $\theta_{(\mathrm{i}-1)}$ represents the parameter vector from the previous iteration.

3.1.7 The proposed CNN network for classifying melanoma in dermoscopic images

Figure 2 illustrates the structure and configuration of the proposed network, consisting of 27 layers. Operating on skin images with a resolution of 128×128 pixels, the network incorporates three hidden layers. The number of output channels in each hidden layer is derived by convolving the image using 3×3 filters, resulting in respective values of 8, 16, 32, 64, 128, and 256. At each layer, the activation process involves the application of ReLU activation and Max pooling, which serves to downsize the image.

The output of the fully connected layer undergoes normalization through the softmax activation function. This function generates positive values that sum up to one, providing classification probabilities used by the classification layer for categorizing inputs into Melanoma and nevi, and for calculating loss.

2.png

Figure 2. The architecture of the proposed CNN network for classifying melanoma in dermoscopic images. It should be noted that this CNN network includes 27 layers

The CNN network proposed in this study is trained using melanoma images obtained from the SIC 2019 and ISIC 2020 challenge datasets. Detailed information about the training and testing images will be expounded upon in Section 5.1.

The training process of the network is facilitated by utilizing the SGD optimizer with momentum, where a predefined number of epochs is set at 50. To improve training effectiveness, the learning rate undergoes reduction by a factor of 0.2 after every 5 epochs. It's important to note that all experiments conducted for this research have been performed using MATLAB.

3.2 Deep transfer learning-based approach

In this study, we utilize transfer learning based on robust pre-trained CNNs to classify melanoma in dermoscopic images. Our approach involves employing ten pre-trained CNN architectures, each renowned for its specific characteristics and advantages: SqueezeNet, GoogLeNet, AlexNet, ResNet-18, ResNet-50, MobileNet-v2, ShuffleNet, NASNet-Mobile, EfficientNetB0, and VGG-19.

The selection of these ten pretrained CNN models in this study was made after careful consideration of their individual strengths and weaknesses, aligning them with the goals of classifying skin cancer in dermoscopic images. Each of these pretrained models has been widely employed in various applications. Models such as MobileNet and EfficientNet are recognized for their highly efficient use of computational resources. On the other hand, models like ResNet and NASNet-Mobile excel in capturing deep features, potentially improving melanoma classification accuracy.

However, some of the pretrained models mentioned above come with limitations. For instance, VGGNet, GoogleNet, and ResNet demand substantial computational resources, limiting their applicability in resource-constrained environments. Additionally, pretrained models like GoogleNet and ResNet possess complex architectures that can lead to longer training times and convergence challenges.

In the subsequent section, we delve into the rationale behind the selection of these models and provide a more detailed analysis of their strengths and weaknesses for the specific task of melanoma classification in dermoscopic images.

3.2.1 SqueezeNet

SqueezeNet, a formidable architecture, comprises fifteen layers, which include two convolution layers, three max-pooling layers, eight fire layers, one global average pooling layer, and a final output layer with softmax activation [31]. Operating with an input size of 227 × 227 RGB channels, SqueezeNet incorporates max-pooling after convolution to generalize input images. Its convolution layers employ 3 × 3 kernels, employing an element-wise activation function that sets all values less than zero to zero [28]. SqueezeNet further leverages fire layers, each encompassing squeeze and expansion stages, maintaining consistent input and output tensor scales.

The inclusion of SqueezeNet in this study stems from its lightweight architecture, rendering it suitable for resource-constrained environments. Its efficient parameter usage and ability to sustain high accuracy are significant strengths. Nonetheless, one limitation is its potential to be less effective in capturing complex features compared to deeper networks.

3.2.2 AlexNet

AlexNet, a widely recognized model for classification and pattern recognition, consists of eight layers, including five convolutional and three fully connected layers. In our adaptation, we align input images with the specifications of AlexNet to ensure compatibility with our data, where the number of classes matches the number of inputs in the original model. The model integrates max-pooling layers positioned between the first two convolutional layers, aiding in the reduction of feature map sizes.

Throughout the training process, a low learning rate has been applied, contributing to smaller weight updates [32]. Unlike some of its counterparts, AlexNet is less computationally intensive, making it an appealing choice. As a pioneering model in the field of deep learning, it was selected for its robustness and established performance. Its relatively simpler architecture and fewer layers contribute to its computational efficiency. However, its architecture might be considered shallow compared to more recent and deeper models, potentially constraining its capacity for feature extraction.

3.2.3 VGGNet

The VGGNet, developed by researchers at Oxford University, is recognized for its pyramidal structure [33]. It comprises a series of convolutional layers followed by pooling layers, contributing to its distinctive architecture. VGGNet is known for its adaptability and suitability for benchmarking across various tasks. Its pre-trained models are commonly utilized in diverse applications. However, it's important to note that VGGNet can be computationally demanding, particularly when initiated from scratch.

The pyramidal structure and simplicity of VGGNet render it an excellent choice for benchmarking purposes. It excels in feature extraction and finds widespread application. Nonetheless, its computational intensity becomes apparent, especially when dealing with large datasets and in resource-constrained environments.

3.2.4 GoogleNet

GoogleNet, also recognized as the inception architecture, is focused on efficiently estimating and distributing dense components within the sparse structure of a convolutional network. It specifically addresses the redundancy issue within deep network activations, emphasizing that not all connections between input and output channels need to be present in the network's design. GoogleNet utilizes convolutions of different sizes (5 × 5, 3 × 3, 1 × 1) to capture data and features at varying scales. It also introduces bottleneck convolutional layers (1 × 1), which play a pivotal role in its design [34].

The inception architecture of GoogleNet is structured to capture features at different scales, offering a comprehensive understanding of images. It adeptly employs various kernel sizes for convolution. However, its complexity can pose challenges to training convergence and demand significant computational resources.

3.2.5 ResNet

ResNet, which stands for residual network, is distinguished by its use of residual modules as the fundamental building blocks. These residual modules are organized in a stacked manner to constitute a complete end-to-end network. One of the unique aspects of the ResNet architecture is its incorporation of tens of thousands of residual layers, enhancing its performance and efficacy in network training. In comparison to AlexNet and VGG, ResNet is notably deeper, being 20 and 8 times deeper, respectively [35].

The deep architecture of ResNet allows it to effectively capture intricate patterns and features in images. Its residual modules play a crucial role in mitigating the vanishing gradient problem, enabling the training of exceptionally deep networks. However, a notable drawback is the increased computational cost associated with its depth.

3.2.6 MobileNet

MobileNet is distinctive for its efficiency, making it well-suited for deployment on devices with limited computational resources, such as mobile devices and low-powered computers. Its architecture is characterized by a minimalistic design, which has proven effective for various tasks, including Palmprint Recognition [36]. It employs various convolutional layers and abstraction layers that utilize depthwise convolutions. MobileNet integrates ReLU activation components and residual layers with specific stride values, contributing to its unique design [37].

This model is ideal for resource-constrained devices due to its low computational requirements. It maintains good performance while minimizing memory and power consumption. However, due to its compact architecture, it may not capture fine-grained details as effectively as larger and more complex models.

3.2.7 ShuffleNet

ShuffleNet introduces a novel ShuffleNet unit aimed at optimizing small networks by integrating the channel shuffle function. It commences with a bottleneck unit, utilizing a 3 × 3 depthwise convolution in the residual branch, followed by pointwise group convolution. The goal of this approach is to align the channel dimension between the residual and shortcut paths. The architecture integrates mixed-grouped convolution, point-grouped convolution, and profoundly separable convolution, maintaining accuracy while significantly reducing computational expenses [38].

The distinctive channel shuffle function of ShuffleNet makes it well-suited for small networks. It effectively balances accuracy and computational efficiency. However, it may not perform as strongly as larger models on more complex tasks.

3.2.8 NASNet-Mobile

The Neural Architecture Search (NASNet) is a contemporary CNN architecture developed by GoogleBrain. It utilizes a reinforcement learning search strategy to discover an efficient building block on a small dataset, which is subsequently transferred to a larger dataset, resulting in state-of-the-art performance with a smaller model size and complexity. In contrast to traditional CNN designs, NASNet explores cells that can form a high-performance block, with the internal structure of these cells determined by a Recurrent Neural Network. This approach involves constructing Normal and Reduction cells and iteratively repeating the building block, with the value of 'n' being automatically calculated.

NASNet's automated architecture search generates efficient models tailored to specific datasets, effectively balancing performance and model size, rendering it suitable for various applications. However, it may require significant computational resources during the search process.

3.2.9 EfficientNet

EfficientNet diverges from traditional CNN design approaches by focusing on augmenting network depth, width, and input resolution in the baseline network. It employs a multi-objective neural architecture search to maximize both accuracy and constrained computational resources. The baseline network, EfficientNetB0, is crafted through this process by utilizing a slightly larger version of the mobile inverted bottleneck convolution (MBConv) and scaling it to create the EfficientNet family of models. This innovative methodology facilitates improved accuracy without significantly inflating computational demands [39].

EfficientNet's strategy for scaling network depth, width, and resolution strikes an optimal balance between accuracy and computational efficiency. It excels in both feature extraction and model size. However, fine-tuning may be necessary to tailor it to specific tasks or datasets.

3.3 Deep transfer learning-based approach for melanoma classification

Training a deep CNN model demands access to a vast image repository [40]. Regrettably, datasets comprising a substantial collection of labeled skin lesion images remain limited. Leveraging the capabilities of deep CNNs often necessitates training on expansive datasets like ImageNet, a process fraught with significant challenges. However, we overcame this challenge by employing transfer learning with robust pre-trained CNN models, including SqueezeNet, AlexNet, VGG-16, GoogleNet, ResNet, MobileNet, ShuffleNet, NASNet-Mobile, and EfficientNet, for the task of melanoma classification.

Notably, our approach introduces a two-classification model specifically tailored to differentiate between malignant and benign skin lesions, contrary to CNN models trained on ImageNet, which are designed for distinguishing among 1000 classes. Thus, three critical stages are involved in achieving effective transfer learning for melanoma classification in dermoscopic images:

1. The initial stage involves selecting a pre-trained model, such as SqueezeNet, AlexNet, VGG-16, GoogleNet, ResNet, MobileNet, ShuffleNet, NASNet-Mobile, EfficientNet. A summary of the strengths and limitations of each model is provided in Section 3.2.
2. In the second stage, we replace the model's classification layer with a SoftMax layer, which involves adjusting weights to substitute class-specific attributes with new layers. This step also encompasses the relocation and replacement of feature maps in convolutional layers with class-specific counterparts.
3. The final stage involves deploying fine-tuned layers that are reused and adjusted to suit the requirements of the melanoma classification task.

5.1 Dataset

The dataset was procured from the Kaggle platform, primarily sourced from the International Skin Imaging Collaboration (ISIC), a global initiative dedicated to advancing melanoma diagnosis and hosting the most extensive publicly accessible repository of high-quality dermoscopic images of skin lesions within the ISIC Archive. The acquired dataset comprises images of melanoma (malignant) and nevus (benign) from the ISIC 2019 and ISIC 2020 Challenge Datasets. Notably, the ISIC 2020 dataset associates each image with a unique patient identity. In total, our dataset comprises 11,449 dermoscopic images encompassing 5,106 melanoma cases and 6,343 nevus cases.

These images are in the ".jpg" color format and exhibit variations in pixel sizes, as depicted in Figure 4. Figure 4(a) provides an example of melanoma images, while Figure 4(b) displays a nevus skin lesion.

4.png

Figure 4. Examples of melanoma and nevus skin lesions from the ISIC dataset

In our experiments, we divided the dataset into 75% for training and 25% for testing. The dataset exhibits a semi-balanced distribution, with 3,830 melanoma images and 4,757 nevus images allocated for training, and 1,276 melanoma images and 1,586 nevus images designated for testing.

The decision to utilize a semi-balanced dataset in our experiments is motivated by the need to strike a harmonious balance between mitigating class imbalance issues and optimizing the overall performance of the model. In the context of melanoma classification, there typically exists a substantial class imbalance between malignant (melanoma) and benign (nevus) skin lesion cases. In real-world scenarios, the occurrence of melanoma cases is significantly less frequent compared to benign cases. Consequently, training a model on a severely imbalanced dataset may result in suboptimal performance, as the model might develop a bias towards the majority class (nevus) and encounter challenges in effectively learning patterns associated with the minority class (melanoma).

By adopting a semi-balanced dataset, our goal is to address this class imbalance challenge while still maintaining a representation of real-world conditions. This approach involves stratified sampling to ensure that both melanoma and nevus cases are adequately represented in both the training and validation sets. By doing so, we provide the model with a more balanced exposure to both classes during training, fostering better generalization and improved performance in identifying melanoma cases.

5.2 Results of the proposed CNN model

In this investigation, we systematically trained four distinct configurations of the proposed CNN model, each distinguished by varying numbers of layers: 15, 19, 23, and 27. The primary objective was to ascertain the optimal configuration that would maximize classification accuracy. As detailed in Section 3.1.7, we introduced the comprehensive CNN model for melanoma classification (Configuration 1), comprising 27 layers, as illustrated in Figure 2. For the remaining three configurations (Configuration 2, Configuration 3, and Configuration 4), we progressively removed Convolution, ReLU, Batch Normalization (BN), and Max-Pooling layers from the complete 27-layer CNN network. Specifically, for Configuration 2 (23 layers), we eliminated layers 18-21 (four layers) from the complete network, as depicted in Figure 2. In the case of Configuration 3 (19 layers), we pruned eight layers from the complete network (layers 14-21). Finally, for Configuration 4, we pruned 12 layers from the complete network (layers 10-21).

The performance of these four CNN configurations was meticulously evaluated, and the results are presented in Table 2. Each configuration underwent assessment using fundamental metrics, including accuracy, AUC, recall, precision, and F1-score, to gauge their suitability for the task of skin cancer classification.

Significantly, Configuration 1 emerged as the top performer, showcasing an impressive accuracy rate of 90.85%. Following closely, Configuration 2 achieved an accuracy of 90.74%. While Configurations 3 and 4 exhibited slightly lower accuracies of 90.32% and 89.55%, respectively, it is noteworthy that even the least performing Configuration 4 demonstrated commendable levels of accuracy. Regarding the Area Under the Curve (AUC), both Configuration 1 and Configuration 2 took the lead, registering an AUC of 0.907. This implies that these configurations adeptly discriminated between malignant and benign skin lesions.

Interestingly, Configuration 4, boasting the fewest layers, secured the highest recall rate at 94.83%, indicating its effectiveness in capturing malignant lesions. However, this heightened sensitivity came at the cost of lower precision. Conversely, Configuration 1 achieved a balanced recall rate of 91.80%, while Configurations 2 and 3 achieved recall rates of 91.17% and 89.72%, respectively. Precision, a metric reflecting the ability to correctly classify instances as positive, revealed that Configuration 2 exhibited the highest precision at 92.58%, closely followed by Configuration 3 at 92.04%.

Table 2. The performance of the four configurations of the proposed CNN model, in terms of accuracy (Acc.), AUC, Recall (Re.), Precision (Pre.) and F1-score (F1)

Configuration	Acc.	AUC	Re.	Pre.	F1
C1: 27 layers	90.85	0.907	91.80	91.69	91.75
C2: 23 layers	90.74	0.907	91.17	92.04	91.61
C3: 19 layers	90.32	0.904	89.72	92.58	91.13
C4: 15 layers	89.55	0.889	94.83	87.39	90.96

Figure 5 further illustrates the confusion matrix of these four CNN configurations. Notably, Configuration 4 produced the highest false-positive rate, incorrectly classifying 217 nevus images as melanoma. Conversely, Configuration 3 demonstrated the highest false-negative rate, with 136 melanoma cases misclassified as nevus.

These results underscore Configuration 1, comprising 27 layers, as the most balanced performer, excelling in accuracy and maintaining a harmonious balance between precision and recall. Configuration 2, featuring 23 layers, closely follows Configuration 1 in performance and offers the advantage of reduced complexity. Configuration 3, with 19 layers, emphasizes precision, while Configuration 4, with 15 layers, prioritizes recall. The choice of configuration should be guided by specific application requirements, such as the importance of minimizing false positives or maximizing sensitivity. Nevertheless, all configurations underscore the remarkable potential of the proposed CNN model in skin lesion classification, demonstrating its effectiveness for this critical medical task.

5.png

6.png

Significantly, ShuffleNet achieves the lowest FP rate, erroneously classifying 94 nevus images as melanoma. It's crucial to highlight that the relatively high FN rate of ShuffleNet impacts its overall performance. ShuffleNet's architectural choices may not optimally align with the characteristics of the dataset used. Dermoscopic images exhibit significant variation in texture, color, and lesion size. A conservative model like ShuffleNet might struggle to generalize effectively across this diversity, leading to misclassifications.

In summary, ResNet-50 and AlexNet emerge as the top-performing CNN models for melanoma classification. Both models excel in accuracy, AUC, recall, precision, and F1-score, demonstrating their effectiveness. The architectural designs and depth of these models significantly contribute to their superior performance, enabling them to capture both low-level and high-level features crucial for melanoma detection. On the other hand, SqueezeNet and GoogLeNet, while achieving high recall rates, exhibit lower precision and, consequently, lower F1-scores. This implies that SqueezeNet and GoogLeNet produce a higher rate of false positives, a critical consideration in diagnosing melanoma, where minimizing false positives is essential. The choice of CNN architecture proves pivotal in obtaining precise and reliable melanoma classification results, with ResNet-50 and AlexNet exhibiting strong suitability for the melanoma classification task.

5.4 Comparisons

In Table 4, we present a thorough comparison between our optimized CNN model's best configuration and the deep transfer learning approach utilizing ResNet-50 pretrained on ImageNet. This comparative analysis illuminates the strengths and trade-offs inherent in each approach.

Table 4. Comparing the proposed CNN model with the deep transfer learning approach based on ResNet-50, in terms of accuracy (Acc.), AUC, Recall (Re.), Precision (Pre.) and F1-score (F1)

ResNet-50 achieves an impressive accuracy of 92.98%, underscoring its robust classification capabilities. This accuracy is approximately 2 percentage points higher than that of our proposed CNN model, which attains 90.85%. It is noteworthy that although there exists a discrepancy in accuracy, the margin is not substantially large, especially considering the significant contrast in complexity between the two models.

Importantly, in terms of the recall rate, ResNet-50 exhibits superior performance compared to our CNN model, indicating its enhanced ability to correctly identify melanoma cases. However, both methods exhibit similar precision rates, emphasizing their capability to accurately classify melanoma cases while minimizing false positives.

This comparative analysis establishes that our proposed CNN model yields result on par with the transfer learning approach employing ResNet-50. What distinguishes our CNN model is its notably lower complexity and lighter computational cost. These attributes render it particularly well-suited for resource-limited devices or environments where computational resources are constrained. Consequently, the choice between the two approaches should hinge on the specific requirements of the application at hand.

Additionally, we conducted a thorough comparison to assess the accuracy of our proposed deep transfer learning-based method against the findings from a previous study. In Table 5, we juxtapose the accuracies achieved by our deep transfer learning approach, utilizing ten pretrained CNN models, with the results obtained by Fraiwan and Faouri [12]. This comparative analysis provides valuable insights into the effectiveness of our approach in relation to existing research.

Notably, Fraiwan and Faouri [12] reported their best accuracy scores with ResNet-18 and ShuffleNet, both achieving an accuracy of 79%. In contrast, our study attains the highest classification accuracy with ResNet-50, reaching an impressive accuracy rate of 92.98%. It's important to highlight that out of the ten pretrained CNN models utilized in our research, three models—AlexNet, NASNet, and VGG-19—were not employed in study of Fraiwan and Faouri [12]. The accuracy rates of all melanoma classification models developed in our study using these ten pretrained CNN models are notably higher, ranging from 5 to 17 percentage points above the results reported by Fraiwan and Faouri [12]. This substantial performance gap underscores the efficacy of the proposed method in enhancing melanoma classification.

Table 5. A comparative analysis of the accuracies achieved by our deep transfer learning approach utilizing ten pretrained CNN models and the accuracies reported by Fraiwan and Faouri [12]