-
Samara S. Santos
Federal Center for Technological Education of Minas Gerais (CEFET-MG), Timóteo, Minas Gerais; Graduate Program in Electrical Engineering, Federal University of Minas Gerais, Av. Antônio Carlos 6627, Belo Horizonte, Minas Gerais; Machine Intelligence and Data Science (Minds) Laboratory, Federal University of Minas Gerais, Av. Antônio Carlos 6627, School of Engineering, Block I, Room 2200, Belo Horizonte, Minas Gerais
-
Tamires M. Rezende
Machine Intelligence and Data Science (Minds) Laboratory, Federal University of Minas Gerais, Av. Antônio Carlos 6627, School of Engineering, Block I, Room 2200, Belo Horizonte, Minas Gerais
-
Marcos A. Alves
Graduate Program in Electrical Engineering, Federal University of Minas Gerais, Av. Antônio Carlos 6627, Belo Horizonte, Minas Gerais; Machine Intelligence and Data Science (Minds) Laboratory, Federal University of Minas Gerais, Av. Antônio Carlos 6627, School of Engineering, Block I, Room 2200, Belo Horizonte, Minas Gerais
-
Frederico G. Guimarães
Graduate Program in Electrical Engineering, Federal University of Minas Gerais, Av. Antônio Carlos 6627, Belo Horizonte, Minas Gerais; Department of Electrical Engineering, Federal University of Minas Gerais (UFMG), Belo Horizonte, Minas Gerais
Keywords:
Skin Cancer Diagnosis, ResNet50, XAI, LIME, Decision Tree, Interpretability
Abstract
Skin cancer is the main type of cancer that affects people all over the world, being melanoma the most feared, due to its rapid spread throughout the body. If it is detected in the early stages, the chances of cures are above 96%. Due to this, approaches to help the clinicians in the correct diagnosis, as well as that focused on the explanations, have been largely explored. In this context, this paper aims to build a binary classification of skin moles model using the ResNet50 and explain its prediction by comparing two known explainer tools LIME and Decision Trees (DT). The ResNet50 architecture achieved results of about 92% in terms of accuracy and up to 91% in sensibility and specificity. When LIME and DT were compared, both showed no fidelity error. However, in terms of stability, measured by the Jaccard index, LIME presented an stability of 0.497 ± 0.473 and DT of 1.0 ± 0.0, showing stability only for the latter. These results were obtained from 30 runs of images randomly chosen from the test base. Through a visual analysis, LIME varied in two of the 5 images from the benign and in one from the malignant lesion. As important as generating good classification models is providing clinicians with good explanation models that are intuitive and consistent.
