[A white matter hyperintensity segmentation model combining 2.5-dimensional spatial context and frequency-domain prior information].
Authors
Affiliations (3)
Affiliations (3)
- College of Electronic and Information Engineering, Shandong University of Science and Technology, Qingdao, Shandong 266590, P. R. China.
- Institute for Digital Medicine and Computer-assisted Surgery, Qingdao University, Qingdao, Shandong 266071, P. R. China.
- Department of Radiology, Affiliated Hospital of Qingdao University, Qingdao, Shandong 265000, P. R. China.
Abstract
Imaging studies indicate that white matter hyperintensities (WMHs) are closely associated with various cerebrovascular and neurodegenerative diseases, and their accurate segmentation is of great significance for the clinical diagnosis and treatment of related brain diseases. However, affected by factors such as insufficient inter-slice resolution caused by magnetic resonance imaging (MRI) voxel anisotropy, blurred boundaries of tiny lesions, and multi-center data heterogeneity, the performance of existing methods still needs to be improved. Targeting 2.5-dimensional (2.5D) MRI slice sequence data, this paper proposes an automatic segmentation network integrating slice context and frequency-domain prior information. First, a inter-slice context Transformer (ISCT) module is designed to extract Z-axis long-range dependencies, in order to reduce inter-slice information loss. Second, a frequency-domain encoder based on learnable discrete wavelet transform (LDWT) is constructed to decouple high- and low-frequency features, and combined with a detail gating module (DGM) to suppress background noise. Finally, a weighted focal Tversky loss function is introduced to enhance the detection robustness for tiny lesions. Experimental results show that the proposed model achieves Dice similarity coefficients (DSC) of 0.800 8 ± 0.077 3 and 0.763 6 ± 0.180 2, and recall of 0.831 1 ± 0.081 8 and 0.852 0 ± 0.204 5 on public and in-house datasets, respectively, and its generalization ability outperforms existing mainstream methods. This model effectively mitigates problems such as lesion omission and cross-domain performance degradation, providing an accurate and robust evaluation tool for the clinical diagnosis and treatment of related diseases.