MSMV-UNet: A 2.5D Stroke Lesion Segmentation Method Based on Multi-slice Feature Fusion

Jingjing Xie, Jixuan Hong, Manjin Sheng, Chenhui Yang

Abstract: Stroke is a severe and life-threatening disease. MRI plays a crucial role in the diagnosis and treatment of stroke, enabling comprehensive analysis of MRI images for accurate localization and qualitative assessment of stroke lesions. A 2.5D stroke lesion segmentation method based on the fusion of multi-slice features is proposed in this study. This method introduces the time information feature between slices in three-dimensional (3D) image data on the basis of a two-dimensional (2D) segmentation model. Multiple encoding paths are densely connected between adjacent slices to utilize the time information feature. To address the problem of difficult segmentation of stroke lesions edges, a Slice-Context Attention Module is proposed to reinforce the differences between adjacent slices in the feature maps. Additionally, considering the multi-perspective features in stroke lesions imaging data, this method proposes to train the segmentation model from three different perspectives: cross section, sagittal plane and coronal plane, and uses soft voting strategy to fuse the results of the three models to form a 2.5D method. Qualitative and quantitative experimental results demonstrate that this method has certain superiority compared to existing methods.