Brain Tumor Segmentation from Multi-Spectral MRI Records: Classical Machine Learning or Convolutional Neural Networks?

Brain tumor segmentation, which involves identifying distinct components like the enhancing core and edema, is a crucial task for ensuring accurate diagnoses and optimizing radiotherapy planning or post-intervention monitoring. However, this process is highly challenging due to the variability in lesion appearances, the potential for noise interference, and differences in MRI scanner sensitivity. This paper represents a comparative study in which XGBoost, the ultimate classical machine learning based method competes with U-net, a well established convolutional network architecture, both being assisted with the adequate preprocessing so that they can achieve their close-to-optimal solution. Training and testing are performed under the very same conditions: a seven-fold cross validation scenario using predefined but random groups of the 259 high-grade glioma records of the BraTS 2019 training dataset. Thorough numerical evaluation suites revealed the superiority of the U-net based solution with respect to a great majority of the statistical accuracy benchmarks. The most relevant quality indicator, namely the average Dice score, is approximately 2% higher for the U-net when segmenting edema, enhancing core or whole tumor, while there is no relevant difference between the two methods in the tumor core segmentation problem.