A Survey on Brain Tumor Segmentation Using MRI Data

Abstract

Gliomas are the most frequent primary brain tumors, with varying degrees of aggressiveness, prognosis, and histological sub-regions, such as peritumoral edematous, necrotic core, active, and non-enhancing core. Variable intensity profiles spread throughout multi-parametric magnetic resonance imaging (mpMRI) images illustrate these sub-regions, representing diverse biological features. In longitudinal scans, the amount of resected tumor is also taken into account while evaluating the apparent tumor for possible progression diagnosis. Furthermore, there is growing evidence that accurate segmentation of multiple tumor sub-regions can provide a foundation for quantitative image analysis to predict patient overall survival. Manual segmentation of brain tumor regions is time-consuming and prone to human error, and its accuracy is determined by pathologists' experience. This study includes about 10 scientific papers that address a wide range of technical topics, including network architecture design, segmentation under imbalanced situations, and multi-modality processes. We use this survey to present a complete assessment of newly established deep learning-based brain tumor segmentation algorithms, taking into account the astonishing breakthroughs produced by state-of-the-art technology.