Transformer Models Advance White Matter Hyperintensities Segmentation

This study compared a 3D convolutional neural network (3D ResNet-50 U-Net) and a Transformer-based model (3D Swin Transformer) for white matter hyperintensities segmentation in brain MRIs. Evaluated on two clinical datasets, the Transformer-based model outperformed the convolutional model across metrics like Dice similarity coefficient, lesion F1-Score, and sensitivity. With limited data and comparable computational resources, the Transformer-based approach demonstrated superior accuracy, making it suitable for clinical applications.