Patch-Based Deep Learning Approaches for Automated 3D Skull Segmentation in CT Imaging

Accurate and efficient skull segmentation from cranial CT scans is a critical prerequisite for computer-aided cranial implant design, surgical planning, and postoperative assessment. In this study, we comparatively evaluate three state-of-the-art deep learning architectures (U-Net, Attention U-Net, and UNETR) for automated 3D skull segmentation. Experiments were conducted on a clinically curated novel dataset of 120 CT volumes encompassing diverse age groups and anatomical variations. Both patched and non-patched inference strategies were investigated to assess their impact on segmentation accuracy and robustness. Quantitative evaluation using a subject-based five-fold cross-validation protocol revealed that patched inference consistently outperformed non-patched processing, with Attention U-Net achieving the highest mean Dice Similarity Coefficient. Qualitative analysis demonstrated that patched inference produced anatomically complete and visually coherent skull reconstructions, particularly in challenging pediatric cases. The proposed approach shows strong potential for integration into clinical workflows, providing high-fidelity 3D reconstructions suitable for patient-specific cranial implant manufacturing.