Inteligência Artificial Explicável: Classificação de Lesões da Cavidade Oral com Modelos de Aprendizagem Profunda

Abstract

The automated histological classification of oral dysplasias represents a significant challenge in digital pathology, demanding deep learning architectures that combine high performance with clinical interpretability. This work aimed to investigate and compare the application of two modern neural network architectures, Vision Transformer (ViT-Base16) and ResNeSt-50, for the automated classification of histological images of oral dysplasia, with a focus on explainability through eXplainable Artificial Intelligence (XAI) techniques. The experimental methodology employed two datasets of murine oral tissue histological images: the first composed of four classes—healthy, mild, moderate, and severe dysplasia—and the second including two genetic variants, wild-type (WT) and knockout (KO), each comprising seven histopathological classes. Initial baseline experiments were conducted, followed by optimizations through histology-specific data augmentation (nine techniques), advanced regularization, and hyperparameter tuning. For explainability analysis, six distinct XAI techniques were implemented: three specific to ViT (traditional Grad-CAM, native attention maps, and enhanced Grad-CAM) and three for ResNeSt (specialized Grad-CAM, CNN attention, and input-based Grad-CAM). Baseline results revealed an initial superiority of ResNeSt (72.41% vs. 64.00% mean accuracy); however, after optimization, this hierarchy was completely reversed, with ViT outperforming ResNeSt (80.79% vs. 78.63%). ViT demonstrated greater responsiveness to optimization techniques (+16.79% average gain vs. +6.22% for ResNeSt), achieving a best accuracy of 87.50% on the four-class dataset. The optimization procedures provided substantial improvements in overall performance (+11.51%), a significant reduction in overfitting (training–validation gap reduced from 37.18% to 19.82%), and excellent validation–test agreement (R² = 0.9988). XAI analysis revealed consistent temporal evolution of explanations and correspondence between model attention and relevant histopathological features, with ViT showing a global focus while ResNeSt emphasized local details. Clinical validation of the explanations confirmed correct identification of stromal invasion regions in carcinomas, epithelial alterations in dysplasias, and tissue thickening in hyperplasias. It is concluded that ViT, when properly optimized, surpasses specialized convolutional architectures in histological classification, offering both high performance and clinical interpretability, thus establishing a foundation for the clinical translation of explainable artificial intelligence in oral pathology applications.