SkinFusionNet: Multi-backbone spatial fusion for robust and explainable melanoma classification

Automated melanoma detection from dermoscopic images remains a critical yet unresolved challenge due to substantial dataset heterogeneity, artifact contamination, and limited model generalization across clinical sources. Existing deep learning methods commonly rely on single-dataset training, single-backbone architectures, or post-hoc interpretability, which restrict robustness and clinical trustworthiness. In this work, we present a novel multi-backbone spatial fusion framework that integrates DenseNet121, EfficientNetV2-B1, and ResNet152V2 to learn complementary mid-level representations from heterogeneous data. The proposed approach uniquely performs spatial feature fusion rather than prediction-level ensembling, enabling richer lesion-aware feature integration. A domain-informed preprocessing pipeline, combining advanced hair removal, illumination correction in LAB space, and adaptive contrast enhancement, further reduces artifact-induced bias. The model is trained on a large merged cohort of approximately 30,000 images from five public datasets using focal loss and evaluated through bootstrap validation. Comprehensive explainability analysis using Grad-CAM++ and Score-CAM confirms consistent lesion-focused attention. The proposed system achieves 94.46% test accuracy with stable performance across bootstrap resamples (95% CI: 93.66%-95.30%), demonstrating enhanced robustness, reliability, and clinical applicability for automated melanoma classification.