An optimal graph convolutional vision neural network with explainable feature optimization for improved skin cancer detection

Despite advancements in skin cancer diagnosis procedures, misclassification rates in early detection remain high, leading to delayed treatments and reduced survival rates. Existing manual diagnostic methods are often prone to inter-observer variability and human error, while traditional machine learning models struggle with imbalanced datasets and insufficient feature generalization. To address these challenges, this work proposes an Optimal Skin Cancer Classification Network (OSCC-Net), developed on the International Skin Imaging Collaboration-2019 (ISIC-2019) dataset. The model integrates an Adaptive Minority Over-Sampling Procedure (AMOP) to balance under-represented lesion classes, ensuring robust learning for minority lesion classes. The Stochastic Neighbourhood T-Distilling driven Score-Weighted Class Activation Mapping (STND-SWCAM) framework is introduced for feature analysis. It performs fine-grained lesion localization and interpretability, enabling better understanding of decisions. In the feature selection stage, a Grizzly Bear Fat Increase Optimizer with Density-Based Spatial Neighbourhood Discovery Algorithm (GBFIO-DSNDA) is employed to enhance discriminative feature extraction by eliminating redundant and noisy features. Finally, classification is performed using a Graph Convolutional Vision Neural Network (GC-VNN), which leverages spatial dependencies among lesion attributes for improved decision-making. Experimental evaluation reveals that, OSCC-Net achieves 98.32% accuracy, 98.43% precision, 98.40% recall, and 98.39% F1-Score, marking a substantial improvement over baselines shown in our experiments.