MVSI-Net: Multi-view attention and multi-scale feature interaction for brain tumor segmentation

Brain tumor segmentation using MRI remains a challenging task due to the high incidence and complexity of gliomas. The irregular variations in tumor location, size, shape, and unclear edge contours of diverse tumor categories contribute to subpar segmentation accuracy. To address these issues, we propose MVSI-Net, a novel MRI brain tumor segmentation method that integrates a multi-view attention mechanism and multi-scale feature interaction into the UNet architecture. Our approach proposes a multi-view attention mechanism that captures global and local features from three different perspectives: channel, content, and position. This mechanism facilitates the localization of the target region and enhances feature representation in lesion areas. Additionally, we design a multi-scale feature interaction module that selectively extracts valuable information from multiple receptive fields of varying sizes, promoting cross-dimensional interaction. As a result, our method enables precise segmentation of the edge contours of different tumor categories. To evaluate the performance of MVSI-Net, we conducted experiments on three widely used datasets: BraTs 2019, BraTs 2020, and BraTs 2021. The experimental results demonstrate that our proposed method outperforms similar approaches in brain tumor segmentation accuracy. In conclusion, our study presents a novel and effective MRI brain tumor segmentation method that addresses the challenges posed by gliomas. However, our model still has certain limitations. Firstly, the model has not been applied in clinical experiments, and there may be challenges in terms of accuracy in certain complex cases. Secondly, further exploration is required to assess the model's generalization capability beyond specific medical image datasets. Moving forward, we plan to address these limitations in future research.