Quantitative magnetic susceptibility imaging (QSM) is an imaging method based on magnetic resonance imaging (MRI) phase signal processing and inversion to obtain tissue magnetic susceptibility distribution, which can generate images reflecting the magnetic characteristics of tissues. QSM reconstruction process is complex, in which dipole inversion stage is the most challenging and decisive link, and traditional methods are easily affected by pathological conditions at this stage, resulting in artifacts and deviations. With the development of deep learning and machine vision technology, using U-network (U-Net) model to improve dipole inversion process can effectively avoid the shortcomings of traditional algorithms. In this paper, the application of the improved model based on U-Net architecture in dipole inversion from 2020 to now is summarized. Firstly, the theoretical concept of QSM is introduced. Secondly, the existing improved models based on U-Net architecture are divided into three categories: improved U-Net based on structural optimization, improved U-Net based on physical constraints and improved U-Net based on improving generalization ability, and their main characteristics and design starting points are sorted out. Finally, the development trend of the future model is prospected and summarized. To sum up, it is expected that the difficulties and challenges of dipole inversion will be solved, the accuracy of QSM images will be improved, and support for disease-aided diagnosis will be provided by summarizing and comparing different improved U-Net models in this paper.
In clinical diagnosis of brain tumors, accurate segmentation based on multimodal magnetic resonance imaging (MRI) is essential for determining tumor type, extent, and spatial boundaries. However, differences in imaging mechanisms, information emphasis, and feature distributions among multimodal MRI data have posed significant challenges for precise tumor modeling and fusion-based segmentation. In recent years, fusion neural networks have provided effective strategies for integrating multimodal information and have become a major research focus in multimodal brain tumor segmentation. This review systematically summarized relevant studies on fusion neural networks for multimodal brain tumor segmentation published since 2019. First, the fundamental concepts of multimodal data fusion and model fusion were introduced. Then, existing methods were categorized into three types according to fusion levels: prediction fusion models, feature fusion models, and stage fusion models, and their structural characteristics and segmentation performance were comparatively analyzed. Finally, current limitations were discussed, and potential development trends of fusion neural networks for multimodal MRI brain tumor segmentation were summarized. This review aims to provide references for the design and optimization of future multimodal brain tumor segmentation models.