Toward more Robust Skull Striping using Custom transformer and Combined Hausdorff loss

Skull stripping is a crucial pre-processing step in medical imaging analysis and has been widely studied using various techniques. Deep learning networks have proven to be effective in many image segmentation tasks, including skull stripping. This study compares four popular deep learning networks, namely U-Net, SegNet, PSPNet, DeepLab, and a novel Vision Transformer for skull stripping. The performance of each network was evaluated on the skull stripping dataset, and the results were analyzed based on various metrics. We also propose a novel loss function to improve model performances. The new loss function uses a combination of shape-aware loss and regional losses to improve the performance of the vision transformer. The study found that Vision Transformer performed significantly better in terms of correctly classifying boundary pixels than non-boundary pixels than the state-of-the-art networks when trained on the new loss function metrics. Furthermore, these findings provide valuable insights into the performance of different deep-learning networks for skull stripping and can aid in the selection of an appropriate network.