Análise Comparativa de Técnicas de Aprendizado de Máquina para Reconhecimento de Células Sanguíneas

Abstract

Artificial intelligence (AI) has established itself as an essential tool in various sectors, including medicine, where it offers innovative and efficient solutions. However, more traditional machine learning approaches still possess vast unexplored potential. This study focuses on the automated classification of four blood cell subtypes (Lymphocytes, Neutrophils, Monocytes, and Eosinophils), a crucial task for medical diagnostics that, when performed manually, can be time-consuming and prone to variability. The study presents a comparative evaluation of machine learning techniques, employing a multimodal approach that combines different types of features manually extracted from images using unsupervised techniques. Specifically, Local Binary Patterns (LBP), histograms of the Hue, Saturation, and Value (HSV) color space, and texture analysis indicators were employed. These feature sets fed the Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Multi-Layer Perceptron (MLP) classifiers. After feature extraction and hyperparameter tuning using K-Fold cross-validation, model performance was evaluated based on validation accuracy. From this, the three best-performing feature combinations were selected for each classifier, allowing for a more objective analysis focused on generalization potential. The results demonstrated that KNN, Random Forest, and especially the MLP model achieved high classification performance. Feature combinations integrating LBP and HSV – particularly LBP+HSV and LBP+HSV+Texture – led to the highest validation accuracies, often exceeding 99%. Notably, the MLP model achieved a validation accuracy of 99.0% when using the combination of all features investigated. This study also outlines promising directions for future research, aiming to broaden the application of this technology in the healthcare field.