Melanoma is the deadliest kind of skin cancer, where premature and accurate recognition considerably improves survival percentages. Conventional diagnostic methods like dermoscopy and biopsy are subjective, time-intensive, and often inaccessible in environments with limited resources. This research introduces an automated melanoma classification system employing Convolutional Neural Networks (CNNs) to address these limitations. Using publicly available datasets, including the Kaggle archive and the International Skin Imaging Collaboration (ISIC), the system utilizes preprocessing methods like resizing, normalization and data augmentation are employed to improve the robustness of the model. And ensure generalization across a wide range of dermoscopic images. The CNN architecture integrates convolutional layers for feature extracting features, pooling for reducing dimensionality reduction, also dropout layers to mitigate overfitting. The model, utilizing categorical cross-entropy loss during training and optimized with the Adam optimizer, achieves an accuracy of 92%, sensitivity of 90%, specificity of 94%, F1-score of 91.2%, and ROC-AUC of 0.96. Interpretability is improved using Gradient-weighted Class Activation Mapping (Grad-CAM), fostering greater clinical trust and user-friendliness. A interface designed to be user-friendly using Streamlit enables real-time classification via image uploads. The system’s affordability and scalability for web and mobile deployment make it suitable for resource-limited settings. Future work includes integrating multimodal data, generating synthetic datasets, and supporting 3D imaging to improve performance and applicability, offering a robust, accessible, and efficient solution for melanoma detection.