A secure and explainable hybrid CNN-transformer framework for trustworthy skin lesion classification

High diagnostic accountability and data integrity in digital health systems are a prerequisite for accurate early-stage classification of malignant skin lesions. Although deep learning models are remarkably successful in laboratory settings, their clinical deployment is limited due to their opaque "black-box" nature and susceptibility to manipulation of data. To that end, we introduce a robust "Trustworthy AI" framework that combines high-performance deep learning ensembles with a private permissioned blockchain to provide an immutable audit trail for automated dermatological screening. To directly overcome hair-induced occlusions and low-contrast artifacts that commonly degrade dermoscopic image quality, our multi-stage preprocessing pipeline applies black-hat morphological filtering for hair removal and Contrast-Limited Adaptive Histogram Equalization (CLAHE) for illumination correction. The core analytic engine is a multi-architecture soft voting ensemble of a MaxViT-ViT hybrid backbone (augmented with scSE and ECA-Net modules), DenseNet-201, and EfficientNet-B4. We further optimize feature calibration with Spatial and Channel Squeeze-and-Excitation (scSE) modules and learnable Generalized Mean (GeM) pooling to better capture fine-grained pathological nuances. The framework achieved an accuracy of 98.95% and precision of 98.05% on a stratified held-out test set drawn from the same combined HAM10000 and ISIC 2019 training distribution, achieving state-of-the-art performance relative to standard baselines; evaluation on a fully independent, out-of-distribution dataset was not performed in this study and is identified as a direction for future work. Finally, we bridge neural computation and clinical practice by employing Grad-CAM and SHAP interpretability analysis to demonstrate that the model attends to clinically relevant features rather than confounding artifacts. This integrated approach - a multi-axis attention hybrid ensemble with dual explainability and blockchain-anchored auditability - offers a secured, explainable template for high-precision computer-aided diagnosis.