Redes Neurais Convolucionais 3D para Classificação da Escala Fazekas em Ressonâncias Magnéticas Cerebrais

Abstract

Neurological disorders such as stroke and dementia increasingly affect aging populations, with White Matter Hyperintensities (WMHs) serving as important imaging biomarkers. The Fazekas scale, commonly used to grade WMHs in brain MRI, is limited by subjectivity and low reproducibility. This study proposes an automated classification approach using 3D brain MRI and deep convolutional neural networks (DenseNet-121 and DenseNet-169). Both models were adapted to process volumetric data and trained on a clinically acquired, radiologist-labeled dataset, following standard preprocessing and augmentation. The 3D DenseNet-121 achieved 89.8% accuracy and 89.2% F1-score in distinguishing mild (Fazekas 0/1) from moderate/severe cases (Fazekas 2/3), while DenseNet-169 reached 87.5% accuracy and 85.3% F1-score. Results in multiclass settings were also competitive. These findings support the use of deep learning for reliable and consistent automated Fazekas grading, contributing to improved neurological diagnosis and early risk assessment.

Ähnliche Arbeiten