This paper introduces a novel approach for skin lesion classification, focusing on melanoma detection, by combining advanced image preprocessing, feature extraction, and Vision Transformers (ViTs). The methodology starts with an enhanced preprocessing pipeline that uses Black-Hat filtering, CLAHE, and Non-Local Means Denoising to improve image quality, reduce noise, and standardize images. Lesion segmentation is performed using Otsu’s thresholding and morphological operations to isolate lesions for more accurate feature extraction. EfficientNetB0, a pre-trained deep learning model, is used for feature extraction, followed by LASSO-based feature selection to identify key features while reducing dimensionality. The final classification step employs Vision Transformers (ViTs), which use self-attention mechanisms to capture global patterns for distinguishing different lesion types. The approach is evaluated on the HAM10000 dataset and compared with traditional models such as Naive Bayes, SVM, Random Forest, Decision Trees, and CNNs. The proposed method outperforms all others, achieving 98% accuracy and minimal misclassifications, particularly in melanoma detection. The research demonstrates that combining ViTs with advanced preprocessing and feature selection techniques offers a promising tool for early melanoma detection, highlighting potential for practical use in dermatology. Further exploration of hybrid models and dataset expansion is suggested to address challenges in detecting visually similar lesions, improving model robustness and generalization. Keywords: Skin Lesion Classification, Melanoma Detection, Vision Transformer, HAM10000 Dataset, Image Preprocessing, Feature Extraction, Lesion Segmentation.