Classification of Skin Lesions Using Machine Learning from Extracted Image Features

The present work addresses the problem of skin lesions, which can appear in different forms and locations on the body and are classified according to their type. Early segmentation and classification of skin lesion images can support a timely diagnosis and appropriate treatment, contributing to better patient recovery. Although several algorithms have been developed for segmenting and classifying these lesions, many still face limitations in accurately extracting the boundaries of the lesion and achieving high classification performance. To improve the reliability of these processes, this study proposes the use of classification models based on features extracted from lesion images obtained from different datasets. The models evaluated are the Support Vector Classifier (SVC) and the Random Forest (RF). Experimental results using the ISIC 2019 dataset show that the SVC model achieves up to 88% accuracy for a 3-class classification task and 79% for 5 classes, while the RF model reaches 91% and 96% accuracy for the same respective tasks. These findings suggest that traditional feature extraction-based machine learning approaches can be effective for multiclass skin lesion classification and potentially suitable for deployment on resource-limited embedded medical devices.