Multiscale Attention-Enhanced U-Net for Robust Skull Stripping in 2D Brain CT Images: A Hybrid Deep Learning and Rule-Based Approach

Skull stripping is an essential preprocessing task in neuroimaging, enabling accurate analysis and intervention of brain tissues by removing non-brain structures from CT images. Traditional methods often suffer with 2D slice-based CT data because of inconsistency in intensity, anatomical distortion, and partial volume effects. Combining a multiscale attention-enhanced 2D U-Net architecture with an anatomically-guided loss function for precise skull segmentation, this article proposes a novel model for brain skull segmentation. The model combines advanced features such as residual connections, attention mechanisms, and hybrid postprocessing techniques to achieve a better accuracy in both global and local anatomical regions. Adversarial domain adaptation also ensures the model generalizes well across diverse datasets. Evaluations demonstrate that the proposed model outperforms existing skull removing methods and provides a robust and adaptable solution suitable for clinical and research applications in neuroimaging.