Hybrid CNN–GNN for skin lesion topology analysis: hybrid CNN–GCN framework for binary dermoscopic skin lesion classification

Abstract Skin cancer is a major global health concern, and early identification of malignant skin lesions is essential for effective clinical assessment and screening. Although Convolutional Neural Networks (CNNs) have achieved remarkable success in dermoscopic image classification, conventional CNN-based methods generally process each image independently and do not explicitly model relationships among lesion representations. This study proposes a hybrid CNN–Graph Convolutional Network (CNN–GCN) framework for binary benign–malignant skin lesion classification using the ISIC 2019 dermoscopic image dataset. Dermoscopic images are first preprocessed for artifact removal, contrast enhancement, and noise reduction. A frozen ImageNet-pretrained ResNet50 extracts 2048-dimensional feature embeddings, which are represented as graph nodes. The novelty of the proposed framework lies in its leakage-controlled investigation of inter-sample relational learning, where graph construction is performed exclusively using training samples and unseen images are evaluated through an independent query-based inference strategy. A sparse cosine-similarity k-nearest-neighbor graph models feature-space relationships, and a two-layer GCN performs neighborhood aggregation to exploit complementary inter-sample information. The proposed framework achieved ,,,98.38% classification accuracy, 97.59% precision, 97.76% sensitivity, and a 97.68% F1-score for malignant lesion detection. These results demonstrate that controlled inter-sample graph reasoning can complement CNN-derived visual features and support computer-assisted dermoscopic lesion classification, while further external and clinical validation remains necessary.