Benchmarking Deep Learning Models for Dermatological Image Analysis: EfficientNet Takes the Lead

Abstract

Skin cancer that spreads quickly and is deadly is called melanoma. If skin cancer is not treated in its early stages, the mortality rate is very high, but when it is correctly identified in its early stages, patients' lives can be saved. With an accurate and fast diagnosis, the patient's chance of survival can be increased. A computer- aided diagnostic support system needs to be created. In this study, Dense201, DarkNet19, and EfficientNet offer 3 different deep transfer learning models for melanoma classification. In addition, an ablation study was conducted in terms of the filter size used in transfer learning. To look at the effect of the filter size, different filter sizes were created in each model and the results were obtained. The ISIC dataset containing 1792 benign and 1464 malignant images was used in the study. According to this study, DenseNet201 provided accurate and reliable results at different filter sizes regardless of their size. Therefore, it is recommended to use DenseNet201 in studies involving the classification of skin lesions.