Review of Multimodal MRI Brain Tumor Segmentation Methods in the Absence of Modality

Abstract

Nowadays, the global prevalence of brain tumors has been increasing steadily. Magnetic resonance imaging (MRI) is widely used in clinical practice to detect brain abnormalities owing to its noninvasive nature. By acquiring multiple pulse sequences, MRI enables tissue characterization with complementary contrasts from a single examination and provides imaging data for subsequent diagnosis. However, prolonged scan time, patient motion and hardware limitations may result in the absence of one or more MRI sequences in practice thus affecting the diagnostic accuracy. Accurate brain tumor segmentation under missing modality condition is stillamazingly challenging in medical image analysis. The review focuses on multimodal brain MRI: T1-weighted, T2-weighted andfluid attenuated inversion recovery (FLAIR) sequences as well as automated brainlesion segmentation. We organize and compare thestate-of-the-art methods for missing modality in a systematic way includingmodality synthesis, shared latent-space learning, knowledge distillation,mutual-information-based learning, Transformer-based fusion and Markov chainbased state-space models. The paper finds that complete-modality inputs provide an important foundation for stable segmentation; the absence of T1-weighted imaging may impair anatomical localization while the absence of T2-weighted or FLAIR imaging may impair the delineation of edema and the whole-tumor region.