CancerSkinNet: A Cross-Attention Multimodal Hybrid Framework Integrating Dual Convolutional Networks and Interpretable Ensemble Classifiers for Early, Equitable, and Explainable Skin Cancer Diagnosis

Deep learning classifiers for dermoscopic images perform well on benchmarks but suffer from poor interpretability, lower accuracy on darker skin and ignore patient metadata. We propose CancerSkinNet, a framework to fill these three gaps in a single architecture with an associated predefined evaluation protocol. ISIC 2020, HAM10000 and Derm7pt images are harmonized to 9 diagnostic classes with patient grouped splits. The feature maps of ResNet-50 and Xception are fused with a 64-dimensional embedding of the metadata through cross attention. The final decision is made by a soft-voting ensemble of a support vector machine, a random forest and XGBoost, and it also receives interpretable ABCD lesion descriptors. The individual typology angle is used to calculate an estimate of the skin tone. SMOTE by class is performed in the feature space. GradCAM and SHAP explanations are assessed by 3 dermatologists. Every component and the statistical plan are built by us: McNemar & DeLong tests with Bonferroni correction, worst-group & maximum-gap equity metrics, and Fleiss κ. The power analysis indicates that 1800 test images are needed to detect a two-point paired accuracy difference, whereas subgroups below 200 images cannot resolve gaps smaller than six points.The framework and protocol describe how claims of accuracy, equity and interpretability should be tested before clinical use.