Malignant melanoma is the worst type of skin cancer and one of the world's fastest-growing malignancies. Early detection and identification of melanoma are critical for a high rate of full cure. Moreover, the importance of patient and dermatologist understanding of early melanoma signs and symptoms constitute a key factor for gaining useful experience concerning the disease. Towards these pillars, automated melanoma diagnosing systems that detect malignant skin lesions at early stages can support dermatologists to obtain objective quantitative markers, to relieve the excessive load of work and facilitate telemedicine capabilities. Deep learning-based approaches’ efficiency has soared in recent years, and they now appear to outperform traditional machine learning methods in classification tasks. In this master's thesis, Convolutional Neural Networks, handcrafted techniques, and metadata are used to extract features from a set of 58,457 dermoscopy skin lesion images. The extracted features from each technique, separately and in combination, are used to train machine learning classifiers towards the creation of a classifier that returns efficient results in terms of accuracy while, simultaneously, exploiting much simpler architectures than the state of the art. A curated ablation study assists in the determination of the base model components for the creation of the final architecture. The proposed method was tested in the SIIM-ISIC 2020 melanoma classification, and it involves the use of a combination of EfficientNet-B0 features, GLCM, LBP, color moments features, and metadata features to improve model performance by 93,97% AUC score.