[Paper Review] Surface-based 3D Deep Learning Framework for Segmentation of Intracranial Aneurysms from TOF-MRA Images

Xi Yang, Ding Xia+2 more|arXiv (Cornell University)|Jun 29, 2020

Intracranial Aneurysms: Treatment and Complications11 references4 citations

TL;DR

This paper proposes a surface-based 3D deep learning framework for intracranial aneurysm segmentation in time-of-flight MRA images, using semi-automatically generated vessel surface models and PointNet++ for fragment classification, followed by SO-Net for precise surface segmentation. The method achieves a 72% Dice similarity coefficient, significantly outperforming existing pixel-based frameworks (46%).

ABSTRACT

Segmentation of intracranial aneurysms is an important task in medical diagnosis and surgical planning. Volume-based deep learning frameworks have been proposed for this task; however, they are not effective. In this study, we propose a surface-based deep learning framework that achieves higher performance by leveraging human intervention. First, the user semi-automatically generates a surface representation of the principal brain arteries model from time-of-flight magnetic resonance angiography images. The system then samples 3D vessel surface fragments from the entire brain artery model and classifies the surface fragments into those with and without aneurysms using the point-based deep learning network (PointNet++). Next, the system applies surface segmentation (SO-Net) to the surface fragments containing aneurysms. We conduct a head-to-head comparison of segmentation performance by counting voxels between the proposed surface-based framework and existing pixel-based framework, and our framework achieved a much higher dice similarity coefficient score (72%) than the existing one (46%).

Motivation & Objective

• To address the limitations of volume-based deep learning in segmenting intracranial aneurysms from TOF-MRA images.
• To improve segmentation accuracy by leveraging surface representations of brain arteries instead of volumetric voxels.
• To integrate human intervention in generating initial vessel surface models for enhanced model performance.
• To develop a framework that classifies and segments aneurysm-containing surface fragments with higher precision than existing pixel-based approaches.

Proposed method

• The framework begins with user-aided semi-automatic generation of a 3D surface model of principal brain arteries from TOF-MRA images.
• The system samples 3D fragments from the vessel surface model for classification.
• PointNet++ is used to classify each surface fragment as containing or not containing an aneurysm.
• Only fragments classified as aneurysm-positive are processed further using SO-Net for detailed surface segmentation.
• The final segmentation is reconstructed from the segmented surface fragments.
• The approach contrasts with traditional pixel-based frameworks by operating directly on surface geometry rather than volumetric data.

Experimental results

Research questions

• RQ1 Can a surface-based deep learning framework outperform existing volume-based methods in intracranial aneurysm segmentation from TOF-MRA?
• RQ2 How does the integration of human-generated surface models affect segmentation accuracy?
• RQ3 To what extent does using PointNet++ for surface fragment classification improve detection of aneurysm regions?
• RQ4 Does SO-Net provide more accurate surface segmentation of aneurysms compared to voxel-based segmentation networks?
• RQ5 What is the quantitative improvement in Dice similarity coefficient when using the proposed surface-based method?

Key findings

• The proposed surface-based framework achieved a Dice similarity coefficient of 72%, significantly outperforming the existing pixel-based framework's 46%.
• The improvement in Dice score demonstrates that surface-based representation enhances segmentation accuracy for intracranial aneurysms.
• The use of PointNet++ for classifying surface fragments enabled precise identification of aneurysm-containing regions.
• SO-Net effectively segmented the surface fragments, preserving anatomical details critical for surgical planning.
• The framework's reliance on semi-automatically generated surfaces reduced noise and improved model generalization compared to purely volumetric approaches.
• The head-to-head comparison confirms that surface-based deep learning is more effective than traditional pixel-based methods for this task.