Vision transformers in multi-modal brain tumor MRI segmentation: A review

Brain tumors have shown extreme mortality and increasing incidence during recent years, which bring enormous challenges for the timely diagnosis and effective treatment of brain tumors. Concretely, accurate brain tumor segmentation on multi-modal Magnetic Resonance Imaging (MRI) is essential and important since most normal tissues are unresectable in brain tumor surgery. In the past decade, with the explosive development of artificial intelligence technologies, a series of deep learning-based methods are presented for brain tumor segmentation and achieved excellent performance. Among them, vision transformers with non-local receptive fields show superior performance compared with the classical Convolutional Neural Networks (CNNs). In this review, we focus on the representative transformer-based works for brain tumor segmentation proposed in the last three years. Firstly, this review divides these transformer-based methods as the pure transformer methods and the hybrid transformer methods according to their transformer architectures. Then, we summarize the corresponding theoretical innovations, implementation schemes and superiorities to help readers better understand state-of-the-art transformer-based brain tumor segmentation methods. After that, we introduce the most commonly-used Brain Tumor Segmentation (BraTS) datasets, and comprehensively analyze and compare the performance of existing methods through multiple quantitative statistics. Finally, we discuss the current research challenges and describe the future research trends.