A Hybrid Capsule Network for Automatic 3D Mandible Segmentation applied in Virtual Surgical Planning

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• 3D reconstruction

Background:

• Accurate mandible segmentation is crucial for 3D virtual surgical planning but is challenging due to anatomical complexity and imaging artifacts.
• Convolutional Neural Networks (CNNs) have improved segmentation, yet require extensive data and face aggregation issues.
• Existing methods struggle with data efficiency and voxel class imbalance in medical image segmentation.

Purpose of the Study:

• To develop a data-efficient method for accurate automatic mandible segmentation on CT images.
• To address the challenges of data aggregation and voxel class imbalance in medical image segmentation.
• To compare the performance of the proposed method against the established 3D-UNet model.

Main Methods:

• Optimized data-efficient 3D-UCaps model combining capsule networks and CNNs for volumetric CT image segmentation.
• Developed a novel hybrid loss function (weighted focal and margin loss) to manage voxel class imbalance.
• Conducted comparative experiments using the public domain database for computational anatomy (PDDCA).

Main Results:

• The proposed 3D-UCaps method achieved an average Dice coefficient of 90% on the PDDCA dataset.
• The 3D-UNet model achieved an average Dice coefficient of 88% on the same dataset.
• The proposed method demonstrated superior performance and required over 50% fewer parameters than 3D-UNet.

Conclusions:

• The optimized 3D-UCaps approach provides accurate mandible segmentation for 3D virtual surgical planning.
• The novel hybrid loss function effectively handles voxel class imbalance.
• The proposed method is a highly effective and parameter-efficient alternative to 3D-UNet for mandible segmentation.