Abstract:
Skin cancer is a common and lethal disease, and its diagnosis is often made difficult by the quite similar visual appearance between malignant and benign lesions. Artificial intelligence (AI), and in particular deep learning methods, for example, convolutional neural networks (CNNs), seems to be a well-suited tool for developing computer-aided diagnosis (CAD) systems to enhance the accuracy. This paper contributes by comparing two well-known CNN architectures, i.e., thedeep ResNet-152 and the lightweight MobileNetV2, to identify the better approach for automatic skin lesion classification. In this paper, we use a three-class dermoscopic image dataset ('melanoma', 'nevus', and 'seborrheic_keratosis') based on highly imbalanced classes to train and test both models via transfer learning. The performance of the models was evaluated using common metrics like accuracy, precision, and recall. The findings also demonstrated that the light-weight model MobileNetV2 obtained a higher overall accuracy (65%) than that of ResNet-152 (63%). An important observation was the inability of the models to handle the minority (seborrheic_keratosis) class, bringing to notice the effect of data imbalance. This work suggests that an efficient architecture, such as MobileNetV2, can offer an appealing performance trade-off for this diagnostic problem, but class imbalance should be addressed in future studies to construct robust and credible systems.
