Comprehensive Review on MRI-Based Brain Tumor Segmentation: A Comparative Study from 2017 Onwards

Abstract

Brain tumor segmentation has been a challenging and popular research problem in the area of medical imaging and computer-aided diagnosis. In the last few years, especially since 2017, researchers have significantly contributed for solving and enhancing the performance of brain tumor abnormality detection and tumor segmentation from magnetic resonance (MR) images. This paper presents a detailed and intensive review of automated brain disease diagnosis and tumor segmentation methods obtained by investigating numerous recent articles. In the first phase, an extensive literature search is conducted with more than 600 articles from medical image analysis, brain disease diagnosis, and tumor segmentation. Around 50% of articles are removed after initial scanning based on certain criteria, i.e., publication year, number of citations, and bibliographic indexing. A total of 161 relevant articles are finally selected in the second phase based on their performance and novelty of the proposed methods. Furthermore, the selected articles are investigated from the perspectives of methodology and performance. Overall methods exploited for brain disease detection and tumor segmentation are categorised into three broad classes, i.e., conventional methods, machine learning-based methods, and deep learning-based methods. As deep learning-based methods are state-of-the-art for computer-aided diagnosis (CAD) nowadays, we investigated several deep learning models, such as the convolutional neural network (CNN), the generative adversarial network (GAN), the U-Net, etc., along with residual block and attention gate, with respect to their learning mechanisms and hyper-parameter tuning. Methods from each class are rigorously reviewed and summarised by identifying their advantages, disadvantages, dataset, MR modality used, and type of images (2D/3D) processed. The methods are also analysed and compared based on their performance in various measures such as dice similarity coefficient (DSC), sensitivity, positive predictive value (PPV), Specificity, Jaccard Index (JI), Accuracy, Hausdorff distance, and computation time. In this review, the high heterogeneity of articles based on different methodologies is considered in light of the recent progress and development of brain tumor detection and segmentation. During analysis, it has been observed that deep learning-based methods, especially various variants of the U-Net model, outperform other approaches for brain tumor segmentation.