SkinFusion-Net: Cross-Attention Hybrid Deep Learning for Skin Disease Classification

Skin cancer is a major global health concern, and early detection is vital for improving patient survival. However, accurate diagnosis is difficult because skin lesions can look similar, clinical assessments can be subjective, and there is a shortage of dermatology experts. Most current AI-based systems for skin disease classification use single-architecture models, which offer limited performance improvements and lack reliable methods for estimating uncertainty, making them less suitable for clinical use. This paper introduces SkinFusion-Net, a hybrid deep learning method that combines ConvNeXt Base, EfficientNet-B3, and ResNet-50 using a cross-attention based feature fusion mechanism. The model is tested on the HAM10000 dataset, which includes 10,015 dermoscopic images across seven skin disease categories. The experimental results show that SkinFusion-Net achieves better classification accuracy and higher melanoma sensitivity compared to individual models, while maintaining computational efficiency for practical use. The inclusion of uncertainty-aware predictions improves the model's interpretability and reliability. These findings show that SkinFusion-Net can be an useful tool for doctors to diagnose skin problems in both universities and hospitals.