A Convolutional Neural Network Approach for Skin Cancer Detection

Abstract— Skin cancer is among the top ten most common cancers worldwide. Early detection significantly increases the likelihood of successful treatment. In this paper, we propose a CNN-based framework to detect and classify various types of skin cancer from clinical images. Our architecture consists of three convolutional blocks employing the swish activation function, followed by pooling, flattening, and fully connected layers. Experimental evaluation on a benchmark dataset demonstrates the model’s potential, achieving competitive classification accuracy across 7 skin cancer classes.

The proposed model is designed to address the inherent challenges in skin cancer detection by leveraging a deep learning architecture that efficiently extracts discriminative features from dermoscopic images. The architecture comprises three convolutional blocks followed by a dense classifier, as illustrated in Figure 1.

Each convolutional block is engineered to gradually extract low to high-level features. The first block applies 32 filters with a 2×2 kernel, utilizing the swish activation function to promote smooth gradients during training. Swish, being a non-monotonic activation function, has been shown to improve model performance and training stability compared to more traditional functions such as ReLU. Following the convolution, a 2×2 max pooling layer is employed to reduce the spatial dimensions, thereby decreasing the computational burden while retaining the most salient features.

The second block repeats this configuration with 32 filters, reinforcing the extraction of finer details from the input images. By stacking multiple convolutional layers, the network is able to learn increasingly abstract representations that capture subtle differences between various types of skin lesions.

In the third convolutional block, the number of filters is increased to 64 while maintaining the 2×2 kernel size. This block further enhances the network’s capacity to detect complex features by capturing more intricate patterns that are characteristic of malignant and benign lesions alike. The subsequent max pooling operation ensures that only the most significant features are retained for further processing.

Following the convolutional layers, the feature maps are flattened into a one-dimensional vector, serving as the input to the fully connected layers. The dense classifier consists of a 64-unit layer with swish activation, which integrates the extracted features into a high-level representation. A dropout layer with a rate of 0.5 is incorporated at this stage to mitigate overfitting by randomly deactivating neurons during training. This regularization step is particularly important given the high dimensionality of the feature space and the potential variability in clinical image datasets.

The final dense layer consists of 7 units corresponding to the 7 classes of skin lesions under study. A SoftMax activation function is applied to generate normalized class probabilities, facilitating multi-class classification by indicating the likeliho

This article presented a CNN-based approach for the automated detection of skin cancer, leveraging swish activation and a three-block convolutional architecture. The proposed model achieved a classification accuracy of 98%, demonstrating its effectiveness indistinguishing among seven different types of skin lesions. A detailed examination of the confusion matrix indicates that the model maintains high precision and recall across all classes, underscoring its robustness and potential clinical utility. Despite the promising results, future work will focus on expanding the dataset and exploring advanced regularization strategies to further enhance the model’s generalization. Additionally, integrating attention mechanisms or transfer learning approaches may offer additional gains in performance and facilitate wider adoption in clinical settings.