*Result*: Dimensionality reduction for Explainable AI in skin lesion Classification: A sparse autoencoder and Lime-Based framework.

*This research employs novel model that integrates innovative dimensionality reduction approaches with Explainable Artificial Intelligence (XAI) methodologies for improving classification accuracy and provide significant perturbation-based visual explanations. The proposed approach is implemented using ISIC-2024 dataset, which presents a challenging benchmark for dermatological image classification. Initially preprocessing is done utilizing an Adaptive Non-Local Mean (ANLM) Filter to enhance image quality by reducing noise, followed by segmentation via Type-2 Fuzzy C-Means to isolate lesion regions accurately. Pre-training is conducted using VisionGG-19, a hybrid model that encapsulates the merits of VGG-19 and Vision Transformer structures for robust feature representation. The framework incorporates the Local Interpretable Model-Agnostic Explanations (LIME) technique for producing interpretable perturbation visualizations. The high-dimensional perturbation data is then processed using a sparse autoencoder. An innovation in this framework is the creation of a graphical latent space, connecting sparse autoencoder outputs to represent weighted graphical structures. The model achieves significant improvements in interpretability and accuracy by the value of (98.3%). [ABSTRACT FROM AUTHOR]*