Enhanced Melanoma Detection Using a Fine-Tuned EfficientNetV2-L Model on Dermoscopic Images
Parul Nasra
Abstract
Melanoma, a severe type of skin cancer, spreads rapidly and makes detection and treatment rather tricky. Improving patient outcomes requires precise and prompt detection. Deep learning methods have shown potential recently in automated melanoma classification using dermoscopic images. This work uses the Melanoma Cancer Image Dataset to examine the utilization of a pre-trained EfficientNetV2-L model for melanoma classification. Renowned for its scalable and efficient design, EfficientNetV2-L offers modern performance for a variety of image categorization purposes. Through an extensive, annotated dataset of dermoscopic images, the EfficientNetV2-L model was refined via transfer learning to differentiate benign lesions from malignant melanoma. The training strategy included preprocessing, augmentation, and data optimization, aiming to increase generalization and model accuracy. Models were evaluated with respect to accuracy, Precision, Recall, and F1-score. Our results reveal that the revised EfficientNetV2-L model considerably raised the accuracy of melanoma classification compared to other deep learning models and traditional methods. According to the study, the model was accurate, generally at 96% overall. EfficientNetV2-L helps dermatologists in early and accurate melanoma detection, making better clinical decisions and enhancing patient outcomes. The results highlight the necessity of ongoing research in this domain and assist the increasing amount of data validating the utilization of advanced deep learning models in medical image processing.