XAI-SkinNet: Leveraging Model Ensembles and Explainable AI for Skin Lesion Detection

Abstract – Skin cancer is the most prevalent cancer globally, with early diagnosis essential in order to decrease mortality. Dermoscopy is commonly used in diagnosis but is time-consuming and subject to variability when it is done manually. This work presents an automated, ensemble-based deep learning system—XAI- SkinNet—to classify skin lesions on the basis of the ISIC 2019 dataset consisting of more than 25,000 dermoscopic images and patient metadata. The system combines three pre-trained convolutional networks: EfficientNetB7, InceptionV3, and ResNeXt, trained separately and together to achieve maximum classification performance. The system uses a two-stream architecture: one stream using images in isolation and another stream using handcrafted features and patient metadata. Texture and meta-level features are combined using dimension reduction and normalisation techniques with deep features. A weighted voting ensemble framework improves prediction stability. The presented system is intended to serve as a reliable, explainable, and precise means for clinical dermatology diagnosis. Keywords: Skin lesion detection, Explainable AI (XAI), Deep learning, Ensemble learning, EfficientNetB7, InceptionV3, ResNet, Medical image classification.