3-D Brain Reconstruction by Hierarchical Shape-Perception Network From a Single Incomplete Image.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotics

Background:

• 3-D shape reconstruction is critical for minimally invasive and robot-guided surgeries, often relying on limited 2-D data.
• Existing methods do not account for information loss due to intraoperative emergencies like bleeding.
• Accurate 3-D organ shape is vital for surgical navigation in confined and indirect operating environments.

Purpose of the Study:

• To propose a novel hierarchical shape-perception network (HSPN) for low-latency 3-D point cloud (PC) reconstruction from single, incomplete images.
• To address the challenge of incomplete visual information in surgical scenarios.
• To enable spontaneous 3-D shape perception and completion of surgical organs, specifically brains.

Main Methods:

• Developed a hierarchical shape-perception network (HSPN) utilizing a branching predictor and hierarchical attention pipelines.
• Incorporated attention gate blocks (AGBs) to aggregate local geometric features from incomplete PCs and internal reconstruction features.
• Generated initial PCs from incomplete images and subsequently completed them with high fidelity.

Main Results:

• The HSPN successfully reconstructed and completed 3-D point clouds (PCs) from single, incomplete brain images.
• Demonstrated superior performance compared to existing methods in qualitative and quantitative evaluations.
• Achieved high accuracy in 3-D shape perception and completion, validated by Chamfer distance (CD) and PC-to-PC error metrics.

Conclusions:

• The proposed HSPN effectively reconstructs and completes 3-D brain shapes from incomplete visual data, outperforming current approaches.
• This method offers a robust solution for 3-D shape perception in challenging surgical environments with limited information.
• The HSPN has significant potential for enhancing navigation in robot-assisted and minimally invasive surgeries.