Abstract
BACKGROUND: Assessing white matter hyperintensity (WMH) is essential for the diagnosis, treatment, and prognosis of multiple sclerosis (MS) and neuromyelitis optical spectrum disorder (NMOSD). MS and NMOSD present dispersed small lesions alongside larger aggregated lesions that are irregularly shaped, posing challenges for the automatic segmentation of WMH on magnetic resonance images. Furthermore, research on NMOSD brain WMH segmentation is limited due to the rare nature of the disease. This study aims to propose a deep learning method for MS and NMOSD brain WMH segmentation. METHODS: In this study, we propose a 2.5D Fourier Convolutional ResUnet (FrC-ResUnet). It utilizes a spectral encoder to extract global information, enabling accurate segmentation of scattered lesions. Additionally, the model incorporates the selective features module (SFM) and the convolutional block attention module (CBAM) to enhance lesion-background differentiation and outline the lesions distinctly. We evaluated our approach on the MS public and local datasets of MS and NMOSD. RESULTS: Compared to U-Net, ResUNet, FC-DenseNet, AttentionUNet, lesion prediction algorithm (LPA) and Sequence Adaptive Multimodal SEGmentation (SAMSEG), the 2.5D FrC-ResUnet achieved the highest Dice similarity coefficient (DSC) on three different datasets, with values of 0.710, 0.667, and 0.822, respectively. CONCLUSIONS: The 2.5D FrC-ResUnet demonstrates accurate and robust segmentation of NMOSD brain WMH. Meanwhile, the model excels in segmenting MS brain WMH, particularly when confronted with irregularly shaped and dispersed lesions.