Convolutional neural networks learn directly from data and are widely used for image recognition and classification. CNNs have been considered one of the best machine learning algorithms to analyze grid-like structured data, such as images. CNNs have shown exceptional performance in image processing problems and computer vision tasks such as localization and segmentation, classification, and detection [15]. A convolutional neural network typically contains tens or hundreds of layers, each of which can be trained to recognize distinct aspects of an image. The output of each convolved picture is utilized as the input to the following layer after filters are applied while training an image with various resolutions. The filters start with detecting basic features such as brightness and edges and become more complex until they reach features that specifically identify the object [15]. There are several hidden layers between a CNN’s input and output layers. These layers carry out operations that alter the data to learn features specific to the data. Convolution, activation (or ReLU), and pooling are the most used layers. The Conv layer is a Convolutional Network’s core building block that does most of the computational heavy lifting. With convolution, convolutional filters are applied to the input images, activating different aspects of the images. By setting negative values to zero and keeping positive values constant, an activation function facilitates faster and more effective training because only the activated characteristics are carried over to the following layer; this is frequently referred to as activation. Pooling reduces the number of parameters the network needs to learn by conducting nonlinear downsampling on the output. Over tens or hundreds of layers, these operations are repeated while each layer learns to recognize various features. The output for the final classification is provided by a classification layer in the CNN architecture’s top layer [15].

This section overviews the most commonly used CNN architectures in the image classification task. Researchers have proposed deep-learning-based skin cancer detection systems based on these architectures.

Krizhevsky et al. [16] proposed an AlexNet, a CNN network that consists of five convolutions and three fully connected layers [17] as shown in Figure 2. AlexNet was trained using 60 million parameters and solved an overfitting problem using dropout layers. AlexNet achieved top-1 and top-5 error rates of 37.5% and 17.0% on ImageNet LSVRC-2010.

VGG is a convolutional architecture proposed by Karen Simonyan and Andrew Zisserman of the Visual Geometry Group, Oxford University. The VGG-16 [18] model achieved 92.7% top-5 test accuracy on an ImageNet [19] dataset that contains 14 million images of 1000 classes. Simonyan and Zisserman [18] replaced large-size kernel filters with several 3 × 3 kernel-sized filters, achieving better performance than AlexNet. VGG has different variants depending on the number of convolution layers; the most common VGG architectures are VGG-16 and VGG-19. The architecture of VGG-16 is shown in Figure 3. VGG-16 comprised thirteen convolution layers, a max-pooling layer, three fully connected layers, and the output layer.

AlexNet won an ImageNet 2012 competition consisting of eight layers. In deep learning, more layers are added to improve performance and minimize the error rate. Adding more layers results in a vanishing gradient in which the gradient becomes zero and exploding gradient in which the gradient becomes too large. He et al. [20] solved the problem of exploding and vanishing gradients by introducing a concept of skip connections. The skip connection bypasses some levels in between to link layer activations to subsequent layers to make residual blocks stacked together to create a ResNet architecture. The layer causing a problem during training can be skipped avoiding exploding and vanishing gradients, and helps train deep neural networks. ResNet architecture is shown in Figure 4.

Huang et al. [21] presented a densely connected convolution network in which each layer was connected to every other layer as shown in Figure 5. In DenseNet, all previous layers’ feature maps are utilized as inputs for each layer, and that layer’s feature map is used as an input for all layers that come after it. DenseNet architecture helped to solve the vanishing gradient problem and allowed feature reuse. Connecting each layer with every other also helped strengthen feature propagation and decrease the number of parameters required to train a deep neural network.

Howard et al. [22] presented a lightweight network, MobileNet, for mobile applications. MobileNet replaces the 3×33×3 convolution operation in standard CNN with the 3×33×3 depthwise convolution and 1×11×1 pointwise convolution operation; using depthwise separable convolution operation instead of standard convolution operation helps in reducing the number of training parameters. The difference between standard convolution operation and depthwise separable convolution used in MobileNet is shown in Figure 6.

The most commonly used deep-learning models used in skin cancer analysis are listed in Table 2.