Classificações explicáveis para imagens de lesões de pele.

Abstract

The skin is the largest organ of the human body and plays a crucial role in protection against external agents, temperature regulation, and sensory perception. Skin lesions compromise this organ and, in severe cases, may indicate dermatological diseases. The use of Artificial Intelligence (AI) has become increasingly common in assisting the diagnosis and prevention of malignant skin lesions. However, its results have sparked intense ethical and moral discussions, particularly when algorithms reproduce biases that directly influence outcomes. In this context, the adoption of Explainable Artificial Intelligence (XAI) stands out by offering greater transparency and understanding of AI decision-making processes. This article aims to demonstrate the classification of skin lesions using artificial neural networks through the classifiers EfficientNetB0, VGG16, and ResNet50, combined with XAI techniques to identify the most relevant features in lesion classification. The ISIC 2019 dataset was used, and among the tested classifiers, ResNet50 achieved the best performance with an accuracy of 83%. To analyze the results, SHAP, LIME, and Grad-CAM were applied, highlighting the regions of the images that had the greatest influence on classification. It is concluded that the application of XAI proves to be a promising resource for the disease classification process, potentially increasing trust in the results obtained.