Metodologia para a classificação multiclasse de imagens histológicas baseada em inteligência artificial explicável ; Methodology for multiclass classification of histological images based on explainable artificial intelligence

Computer-aided diagnostic (CAD) systems have been studied as a tool to decrease inter-expert variability and streamline the diagnostic process. In this type of system, computer vision and machine learning techniques are usually employed in the automatic diagnosis from histopathological images. Despite resulting in high precision classifiers, many of these methods are black boxes, in which the knowledge used in the decision is implicit in the model, or is represented in a complex way. This aspect decreases the reliability of the system and makes it difficult to debug. In this context, this work proposes a methodology that integrates multiclass classification and explainable artificial intelligence (XAI) methods to build more interpretable and reliable CAD systems. The prediction is made from morphological and non-morphological descriptors, extracted from the cell nuclei identified in the segmented images. These descriptors are pre-processed and used in the construction of the models. In order to improve the understanding of classifications, our methodology integrates different XAI techniques. A strategy based on binary predictors estimates the reliability of model decisions, categorizing them as reliable, uncertain or inconclusive. Then, methods based on Anchors and Shapley Additive Explanations (SHAP) are used to explain the global and local behavior of the model. A new way of displaying anchors provides an alternative to interpret the predictive model's decision. Finally, a data display approach, based on histograms and 3D visualizations, helps in the diagnosis of uncertain or inconclusive cases. This methodology was evaluated in the classification of histological images of non-Hodkin lymphomas and oral dysplasias, resulting in models with an average accuracy of around 94% and 92%, respectively. Based on the interpretation analyses, it was possible to improve the understanding of the behavior of multiclass classifiers. ; CNPq - Conselho Nacional de Desenvolvimento Científico e Tecnológico ; FAPEMIG - Fundação ...