Akademik Veri Yönetim Sistemi
Signal, Image and Video Processing, 2024 (SCI-Expanded)
Automatic classification of dermoscopy images plays a crucial role in the early diagnosis and treatment of serious diseases like skin cancer. However, it poses several challenges, including similar appearance lesions, different types of skin structures, variations in lesion stages, insufficient or inaccurate data, and artifacts present in dermoscopy images. In skin lesion classification tasks, deep learning-based methods have recently demonstrated superior performance compared to traditional machine learning-based methods. In this study, a novel ensemble-based approach is designed for skin lesion classification by leveraging the diverse information captured by different architectures of ConvNeXt models which have been demonstrated to achieve comparable performance to most vision transformers by utilizing a CNN backbone. More specifically, firstly, different versions of pre-trained and fine-tuned ConvNeXt models, namely Tiny, Small, Base, and Large, were used for the classification of skin lesion images to analyze and compare classification performances on the publicly available ISIC 2019 dataset. Among the individual models, ConvNeXt-Large achieved the highest accuracy rate of 97.2%, making it the top-performing model. Then, all four ConvNeXt models were fused using confidence scores to improve classification accuracy. The ensemble approach achieved an overall classification accuracy of 97.7%, surpassing both the performance of individual models and state-of-the-art methods. Additionally, a sensitivity value of 84.2% and a specificity value of 97.9% were obtained. The findings of this study provide evidence that the proposed approach effectively and accurately classifies skin lesions from dermoscopy images.