| Abstract [eng] |
This study addresses the problem of classifying skin lesions. It reviews existing solutions for skin cancer segmentation and classification and proposes an improved solution for melanoma detection from a spatial and spectral analysis of superpixel graphs. The improvement is highlighted by the application of the system to analyze multispectral images and to perform pruning of the extracted features. Since different features play an important role in training a random forest model, thresholds based on MDI and SHAP values were included. For the experiments, a dataset of 472 images obtained from the Lithuanian University of Health Sciences was used, consisting of images of skin lesions extracted with SIAscope. The dataset consists of 5 classes and each sample consists of 5 images: an RGB melanin image, a greyscale melanin image, a hemoglobin image, a collagen image, and a derma melanin image. The classification of skin lesions from digital images is a complex task, but it is of great importance in the medical field in order to speed up the identification of skin lesions and to allow their remote assessment. The highest accuracy achieved was 96.87% for binary classification, compared to 84.37% for five classes. The project solution is detailed through the design section. The Experiments section presents the results and conclusions. |
