MedmeshCNN - Enabling meshcnn for medical surface models.

Area of Science:

• Medical imaging analysis
• Deep learning on 3D meshes
• Computational anatomy

Background:

• Deep learning on 3D meshes, exemplified by MeshCNN, shows promise for analyzing complex anatomical data.
• Existing MeshCNN methods face limitations with the high diversity and fine-grained details characteristic of medical surface models.
• Medical imaging generates abundant complex 3D surface data that could benefit from advanced mesh processing techniques.

Purpose of the Study:

• To adapt and expand the MeshCNN framework for processing complex, diverse, and fine-grained medical surface data.
• To develop MedMeshCNN, a specialized deep learning model for medical mesh analysis.
• To overcome limitations of existing methods in handling patient-specific medical data and imbalanced pathological distributions.

Main Methods:

• MedMeshCNN builds upon MeshCNN, incorporating enhanced memory efficiency for processing patient-specific properties.
• The framework is designed to handle highly imbalanced class distributions common in segmenting pathological medical structures.
• Implementation focuses on direct operation on irregular, non-uniform 3D medical meshes.

Main Results:

• MedMeshCNN achieved 63.24% Intersection over Union (IoU) in segmenting complex intracranial aneurysms and surrounding vessels.
• Segmentation of pathological aneurysms specifically reached an IoU of 71.4%.
• Demonstrated successful application on fine-grained medical surface meshes.

Conclusions:

• MedMeshCNN effectively extends MeshCNN capabilities to complex medical surface meshes.
• The model successfully addresses challenges posed by imbalanced class distributions in pathological findings.
• Patient-specific properties are retained during the processing of medical data using MedMeshCNN.