Hybrid explainable deep learning framework for skin disease detection and future risk prediction using Sakaguchi function with CNN-transformer models

• Skin diseases pose a significant global health challenge, encompassing a wide range of bacterial, fungal, viral, and parasitic conditions. In this research, an intelligent and interpretable deep learning-based system for automated skin disease detection and future risk prediction using dermatoscopic images. • The study initially applies a comprehensive preprocessing pipeline including hair removal, bilateral filtering, K-means clustering for artifact suppression, and histogram equalization to enhance lesion clarity and preserve edge detail. In addition to this, We introduced the Sakaguchi function as an advanced preprocessing enhancement to improve lesion region segmentation and contrast optimization. • After applying both preprocessing techniques, features were extracted using pretrained CNN models like ResNet50 and InceptionV3, followed by the development of hybrid deep learning architectures such as CNN-BiLSTM, CNN-Capsule Networks, and CNN combined with Vision Transformers (ViT). • Upon comparative evaluation, the CNN + ViT model with Sakaguchi-based preprocessing yielded superior performance, achieving an accuracy of 96.10 %, precision of 96.40 %, recall of 95.80 %, and F1-score of 96.10 %, outperforming traditional CNN approaches and preprocessing methods. • Explainable AI techniques including Grad-CAM and Layer-wise Relevance Propagation (LRP) were incorporated to visualize and interpret model decisions, enhancing clinical trust. Furthermore, future disease risk was modeled using LSTM architectures and risk classification via XGBoost. Skin disorders, which include a variety of bacterial, fungal, viral, and parasitic illnesses, represent a major global health concern. Using dermatoscopic pictures, this study suggests an intelligent and comprehensible deep learning-based method for automated skin disease detection and risk prediction. Hair removal, bilateral filtering, K-means clustering for artifact suppression, and histogram equalization to improve lesion clarity are all part of the preprocessing procedure. In addition to this, the Sakaguchi function, mathematically expressed as f ( x ) = x · exp ( − α x 2 ) where α is a tunable contrast-control parameter, is introduced as an advanced enhancement technique. Unlike conventional methods such as CLAHE, it adaptively enhances local contrast while preserving edges and minimizing noise amplification, thereby improving lesion texture visibility and segmentation accuracy. Features were extracted using pretrained CNN models like ResNet50 and InceptionV3, followed by hybrid deep learning architectures such as CNN-BiLSTM, CNN-Capsule Networks, and CNN combined with Vision Transformers (ViT). With an accuracy of 96.10 %, precision of 96.40 %, recall of 95.80 %, and F1-score of 96.10 %, the CNN + ViT model combined with Sakaguchi-based preprocessing produced the best results. Clinical dependability was improved by using explainable AI approaches as Grad-CAM, Layer-wise Relevance Propagation (LRP), and SHAP to visualize and analyze model decisions. GANs were used to generate synthetic data that enhanced performance on minority classes in order to remedy class imbalance. The Overall, the CNN-ViT hybrid framework’s incorporation of the mathematically defined Sakaguchi enhancement shows theoretical and practical improvements in generalization, interpretability, and robustness, making it appropriate for dermatological diagnosis and progression monitoring in the real world.