Convolutional neural network-based skin cancer classification with transfer learning models

Abstract

Skin cancer is a medical condition characterized by abnormal growth of skin cells. This occurs when the DNA within these skin cells becomes damaged. In addition, it is a prevalent form of cancer that can result in fatalities if not identified in its early stages. A skin biopsy is a necessary step in determining the presence of skin cancer. However, this procedure requires time and expertise. In recent times, artificial intelligence and deep learning algorithms have exhibited superior performance compared with humans in visual tasks. This result can be attributed to improved processing capabilities and the availability of vast datasets. Automated classification driven by these advancements has the potential to facilitate the early identification of skin cancer. Traditional diagnostic methods might overlook certain cases, whereas artificial intelligence-powered approaches offer a broader perspective. Transfer learning is a widely used technique in deep learning, involving the use of pre-trained models. These models are extensively implemented in healthcare, especially in diagnosing and studying skin lesions. Similarly, convolutional neural networks (CNNs) have recently established themselves as highly robust autonomous feature extractors that can achieve excellent accuracy in skin cancer detection because of their high potential. The primary goal of this study was to build deep-learning models designed to perform binary classification of skin cancer into benign and malignant categories. The tasks to resolve are as follows: partitioning the database, allocating 80% of the images to the training set, assigning the remaining 20% to the test set, and applying a preprocessing procedure to the images, aiming to optimize their suitability for our analysis. This involved augmenting the dataset and resizing the images to align them with the specific requirements of each model used in our research; finally, building deep learning models to enable them to perform the classification task. The methods used are a CNNs model and two transfer learning models, i.e., Visual Geometry Group 16 (VGG16) and Visual Geometry Group 19 (VGG19). They are applied to dermoscopic images from the International Skin Image Collaboration Archive (ISIC) dataset to classify skin lesions into two classes and to conduct a comparative analysis. Our results indicated that the VGG16 model outperformed the others, achieving an accuracy of 87% and a loss of 38%. Additionally, the VGG16 model demonstrated the best recall, precision, and F1- score. Comparatively, the VGG16 and VGG19 models displayed superior performance in this classification task compared with the CNN model. Conclusions. The significance of this study stems from the fact that deep learning-based clinical decision support systems have proven to be highly beneficial, offering valuable recommendations to dermatologists during their diagnostic procedures.