Performance Comparison of Deep Learning Architectures for Skin Cancer Classification

Abstract

The diagnosis of skin cancer is a serious health problem worldwide. Early diagnosis improves the quality of life and survival of patients. Accurate diagnosis is challenging and requires expertise due to the diversity and similarity of skin lesions. Deep learning techniques are frequently recommended in the literature to solve these challenges and to achieve robust results in the diagnosis process. In this study, the comparative potential of different deep learning models for the classification of skin cancer types is investigated. Within the scope of the study, training and testing processes were performed with ISIC Skin Cancer Image Dataset and six variations of EfficientNet and MobileNet. Six different deep learning models were evaluated: EfficientNetB0, EfficientNet-B1, EfficientNet-B2, MobileNet-V2, MobileNet-V3-Small and MobileNet-V3-Large. As a result of the test processes within the scope of the study, achieving accuracy scores of 86.91 %, 87.35 %, 87.50 %, 85.59 %, 86.76 %, and 89.41%, respectively. In the comparison between EfficientNet and MobileNet models and six variations, MobileNet-V3-Large model was found to be first in classifying skin cancer with 90.59% Recall and 89.53% F1 score. It is predicted that the systems to be developed with the MobileNet-V3-Large model will decrease the diagnosis time of physicians and increase the rate of successful diagnosis. © 2024 IEEE.