Classificação de imagens histológicas utilizando algoritmo polinomial de base hermite e regularizador

Abstract

Diagnosis support systems are crucial for the analysis of histological images, but their performance depends on the extraction of robust features. Although fractal geometry and eXplainable Artificial Intelligence (XAI) descriptors hold relevance, recent studies point to the superiority of features extracted by deep learning. Therefore, this work proposes a classification system for histological images based on the Hermite Polynomial (HP) algorithm, associated with multi-source descriptors (fractal geometry, deep learning, and XAI). To handle the high dimensionality of the data, the system integrates a LASSO regularizer for feature selection. The computational complexity inherent to HP, in turn, is overcome through a parallel implementation that distributes the training and evaluation process of multiple feature subsets among CPU cores. The experimental evaluation, conducted on binary and multi-class databases, demonstrated the superiority of descriptors extracted from the ResNet-50 (RN50) network in a comparative analysis with fractal and XAI-based approaches. With the RN50 descriptors, the methodology achieved a performance with accuracy (ACC) of up to 100% and an imbalanced accuracy measure (IAM) of 1.00 in binary data classification scenarios. In multi-class databases, the ACC value was higher than 98% and the IAM was 0.89. Additionally, in robustness tests against noise, the proposed system maintained its superiority, presenting significantly higher accuracy than other machine learning algorithms even with 50% noise fin the attributes, in both binary and multi-class scenarios. The results consolidate the combination of the HP classifier with regularization and RN50 descriptors as a high-precision and robust approach, with the potential to assist the decision-making of specialists in clinical practice.