Back to the Basics: A Resource-Efficient PCA and Voting Ensemble (KNN, SVM, XGBoost) for Brain Tumor Classification, Challenging Deep Learning Performance (De volta ao básico: uma votação de conjunto (KNN, SVM, XGBoost) com PCA e uso eficiente de recursos para classificação de tumores cerebrais, desafiando o desempenho do aprendizado profundo)

Resumo

https://doi.org/10.5281/zenodo.20215092

The present study addresses the use of artificial intelligence (AI) in the field of neuro-oncology, particularly in the classification of brain tumors using Magnetic Resonance Imaging (MRI). This applied research, based on an exploratory method, adopts an inductive experimental approach to perform a comparative analysis between Deep Learning architectures (DeiT and ResNet) and classical supervised models optimized through Ensemble Learning (Voting Classifier: XGBoost, SVM, and k-NN). The study faced the challenge of domain shift, implementing a preprocessing pipeline with histogram equalization techniques (CLAHE) and a focus on PCA. The results highlighted the nuances of each approach: DeiT achieved high performance, reaching an average of 96% accuracy in generalization, though with high computational cost and low interpretability. In contrast, the Ensemble model achieved an average of 82% accuracy, standing out for its CPU efficiency and greater transparency. The study concludes that, although deep networks define the upper limit of accuracy, optimized classical approaches represent a viable and cost-effective alternative for medical decision support, enabling the selection of the most suitable tool according to infrastructure constraints and ethical requirements.