Lesion Segmentation in Skin Cancer Images using Fusion Model via Deep Learning Networks

Skin cancer, one of the most prevalent and life-threatening cancers globally, has become a focus of deep learning applications due to its significant impact on diagnostic accuracy. This research specifically addresses lesion segmentation in skin cancer images, recognizing its direct influence on classification precision. Six diverse deep learning models, including DeepLabV3+, EfficientNetB7, VGG19, Attention-UNet, MultiRes-UNet, and Transformer-UNet, were implemented. The most effective models, DeepLabV3+ and EfficientNetB7, had their predictions combined using an average fusion method. The model underwent training on the ISIC 2017 dataset and testing on the PH2, ISIC 2016, and ISIC 2018 datasets. Results indicate the proposed model achieved exceptional performance, with an accuracy of 95.9%, 92.8%, 94.5%, Jaccard index of 85.6%, 79.6%, 85.8%, and Dice Coefficient (DC) score of 91.8%, 87.5%, 91.9% on the ISIC 2016, ISIC 2018, and PH2 datasets correspondingly. These scores position the proposed model among the highest compared to other published methods, showcasing its accuracy and effectiveness in skin lesion segmentation.