Automatic Brain White Matter Hyperintensities Segmentation with Swin U-Net

This work proposes an automatic segmentation approach to detect White Matter Hyperintensities (WMH) using Fluid Attenuated Inversion Recovery (FLAIR) and the corresponding T1 weighted MRI images. In this work, we used the Swin U-Net architecture based on Transformers and compared its performance with two currently reported CNN U-Nets architectures. Sixty pairs of images and their corresponding ground truth labels were used from the MICCAI challenge. The following metrics were obtained: Dice similarity score of 0.80; lesion F1 score of 0.63; Hausdorff distance of 3.16 mm; and, 16.93 average volume difference, locating the Swin U-Net second in the ranking of the tested algorithms. However, the computational resources needed to process the imaging data were the lowest compared to the previous U-Nets. Therefore, the Swin U-Net architecture shows potential to become a promising and fast tool for segmenting medical images.