Comparison of CNNs and ViTs for the Detection of Human Skin Lesions

This study aims to evaluate the effectiveness of eight deep-learning models in skin lesion detection using dermatoscopic images. For this purpose, two architectures have been rained and compared: Convolutional Neural Networks (CNN) with models such as EfficientNet B0, B1, B2, and B3 and Vision Transformer (ViT) with the models base 16, base 32, large 16, and large 32. These models were trained using the HAM10000 dataset, a widely recognized database containing diverse dermatoscopic images with labeled clinical diagnoses. The main objective of this research is to evaluate the performance of these models in the classification and differentiation of malignant and benign skin lesions to determine which one might offer superior performance. After training the eight network models, it was determined that the EfficientNet B1 model, after 2000 epochs, achieved the best accuracy with 80%. In second place is the ViT base_patch_32 model with 76.67%. The performance of the ViT model improved significantly by adjusting the learning rate parameters and the number of epochs, achieving 85%. However, Vit required more computational resources to achieve these results than the CNNs.