Automated segmentation of head CT scans for computer-assisted craniomaxillofacial surgery applying a hierarchical patch-based stack of convolutional neural networks

Category	Detail
Condition	Craniomaxillofacial anatomical segmentation in head CT scans
Key Mechanisms	Hierarchical patch-based stack of 3D convolutional neural networks (CNNs) based on U-Net architecture for automated segmentation
Target Population	Patients undergoing computer-assisted craniomaxillofacial surgery, including head and neck cancer and trauma cases
Care Setting	Preoperative planning and intraoperative guidance in surgical and radiological settings

Automated segmentation includes bone structures (viscerocranium, skull base, mandible), foramina/canals, paranasal sinuses, and soft tissues (ocular globe, optic nerve, extraocular muscles).
Patchwork toolbox uses nested patches with fixed matrix size (32x32x32 voxels) and a four-level scale pyramid to balance field-of-view and spatial resolution.
Input CT images are processed with a channel splitting layer separating Hounsfield unit ranges into 11 feature channels, mimicking radiological windowing.

Use thin-slice (<1 mm) head CT scans reconstructed with Br36 or Br40 algorithms for optimal segmentation input.
Resize image resolution to isotropic 1 mm3 voxels for computational efficiency and comparability.

Apply hierarchical patch-based CNNs with multi-scale U-Net architectures for automated segmentation to support surgical planning and navigation.
Employ Adam optimizer with learning rate 0.001 and adapt loss functions for small and large anatomical labels to improve segmentation accuracy.

Compare volumetric and morphological features between preoperative planning and intra-/postoperative imaging to evaluate surgical outcomes.

Potential limitations in segmentation accuracy for very small anatomical structures requiring specialized loss functions.
Hardware capacity may limit sample size and pyramid depth, affecting segmentation resolution and field-of-view.

Patients requiring detailed 3D anatomical segmentation for craniomaxillofacial surgery planning and intraoperative guidance

Automated segmentation facilitates fabrication of patient-specific cutting guides and implants, improving surgical precision and outcome evaluation.

Utilize multi-scale patch-based CNN architectures to capture both global context and fine anatomical details.
Incorporate radiologically inspired channel splitting of CT input data to enhance feature extraction across tissue types.
Adjust loss functions to address class imbalance, especially for small anatomical structures, to improve segmentation performance.

Clinical Scorecard: Hierarchical Patch-Based Convolutional Neural Networks for Automated Segmentation of Head CT Scans in Computer-Assisted Craniomaxillofacial Surgery