Hierarchical Patch-Based CNNs for Automated Segmentation in Craniomaxillofacial Surgery

Overview

This study evaluates a hierarchical patch-based stack of convolutional neural networks (CNNs) based on the U-Net architecture for automated segmentation of 18 craniomaxillofacial structures on head CT scans. The approach leverages multiscale 3D patches and channel splitting of Hounsfield units to improve segmentation accuracy of bone, foramina, sinuses, and soft tissue relevant to computer-assisted surgery.

Background

Computer-assisted craniomaxillofacial surgery relies heavily on accurate three-dimensional segmentation of anatomical structures from CT imaging for preoperative planning, intraoperative guidance, and postoperative evaluation. Traditional manual segmentation is labor-intensive, and while atlas- or model-based methods have been used, recent advances in artificial intelligence, particularly convolutional neural networks like U-Net, have revolutionized medical image segmentation. However, AI-based segmentation specifically tailored for craniomaxillofacial surgery remains underexplored. This study aims to fill this gap by applying a hierarchical patch-based CNN approach to segment multiple relevant anatomical structures automatically.

Data Highlights

The study segmented 18 anatomical structures including viscerocranium, skull base, mandible, foramina/canals, paranasal sinuses, ocular globe, optic nerve, and extraocular muscles. CT scans with slice thickness under 1 mm and specific reconstruction algorithms were used. The network employed a four-level scale pyramid with fixed 32x32x32 voxel patches and isotropic 1 mm resolution at the finest scale. The input CT images were split into 11 feature channels based on Hounsfield unit ranges to mimic radiological windowing. Training used the Adam optimizer with a learning rate of 0.001 and a combination of softmax cross-entropy and top-K binary cross-entropy loss functions tailored for large and small labels respectively.

Key Findings

• The hierarchical patch-based CNN architecture effectively segments both large bone structures and small anatomical features such as foramina and canals.
• Channel splitting of CT input into multiple HU-based feature channels enhances the network's ability to differentiate tissue types relevant for segmentation.
• The multiscale patch approach balances field-of-view and spatial resolution, enabling detailed segmentation within hardware constraints.
• Use of different loss functions for large and small labels improves training focus on small, difficult-to-segment structures.
• The network architecture is based on a modified 3D U-Net with skip connections and transposed convolutions, facilitating efficient learning and reconstruction of segmentation maps.

Clinical Implications

Automated segmentation using this hierarchical patch-based CNN can significantly reduce the time and effort required for preparing 3D models in computer-assisted craniomaxillofacial surgery. Accurate delineation of complex anatomical structures supports improved preoperative planning, fabrication of patient-specific surgical guides and implants, and intraoperative navigation. This approach may enhance surgical precision and patient outcomes by providing reliable and reproducible segmentation results.

Conclusion

The hierarchical patch-based convolutional neural network approach demonstrates promising capability for automated, detailed segmentation of diverse craniomaxillofacial structures on head CT scans, supporting advanced computer-assisted surgical workflows.

References

1. Reisert et al. -- Patchwork Toolbox for Multiscale 3D CNN Segmentation
2. Ronneberger et al. 2015 -- U-Net: Convolutional Networks for Biomedical Image Segmentation
3. Çiçek et al. 2016 -- 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation
4. Kingma & Ba 2014 -- Adam: A Method for Stochastic Optimization
5. Wang et al. 2019 -- Top-K Loss for Training Deep Neural Networks with Noisy Labels