Deep feature fusion for melanoma detection using ResNet-VGG networks with attention mechanism

Skin cancer, especially melanoma, the most severe type, has increased in recent decades. Identifying skin carcinoma infection is important to improve treatment outcomes and patient survival. However, accurate identification of skin lesions remains problematic due to the large variability and heterogeneity of the source material, especially lesion characteristics, even though significant progress has been made in computer vision. To overcome these challenges, this study proposes the channel and spatial attention-based hybrid model (CA-SLNet), which combines the architectures of residual network-50 (ResNet50) and visual geometry group-16 (VGG-16) with updated attention mechanisms. It applies channel attention to enhance the important feature maps and spatial attention to illuminate the important area in the feature. CA-SLNet was evaluated on three dermatological benchmark datasets, ISIC-2016, ISIC-2017, and Ph 2 . The proposed model achieved an accuracy of 98.59%, 96.80%, and 99.50% for the ISIC-2016, ISIC-2017, and Ph 2 datasets and high area under the curve (AUC) values. The proposed model is compared with state-of-the-art methods and emphasizes the robustness and precision of the model in differentiating between benign and malignant lesions. Although slightly more false negatives were observed in the ISIC-2017 dataset, the model’s overall performance proves its suitability for clinical application. CA-SLNet provides a reliable solution for early and accurate melanoma detection by effectively integrating complementary features and minimizing noise and redundancy.