Segmentação Automática de Vértebras em 3D e Diagnóstico de Fraturas em Imagens de Ressonância Magnética Utilizando Dense U-Net e XGBoost

Abstract

Lumbar pain is a common reason for clinical visits and magnetic resonance imaging is frequently used in systems to support the diagnosis of spinal pathologies. Aiming to improve and automate this process, this study proposes the use of computational techniques for the segmentation of vertebrae in magnetic resonance imaging, with the purpose of performing further analysis about pathologies in the spine. To achieve this goal, Deep Learning and Machine Learning architectures are used: Dense U-Net for 3D segmentation and CNN with XGBoost for the classification of vertebrae with rupture or not. The results show that Dense U-Net is promising to localize the vertebra region, obtaining an average Dice Coefficient value of 90.82%, thus overcoming several important studies focused on the problem. Classification was also efficient, with the best results being 96.88% for accuracy and 88.89% for sensitivity.