Pseudo-Rendering for Resolution and Topology-Invariant Cortical Parcellation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Cortical mesh parcellation is crucial for neuroimaging analysis.
• Existing methods often require sphere-like mesh topology, excluding imperfect or noisy data.
• Deep learning struggles with non-descript surface scans and long-range dependencies.

Purpose of the Study:

• To develop a novel cortical mesh parcellation method robust to topological defects and surface scan noise.
• To reframe mesh parcellation as a 2D segmentation task using a rendering-based framework.
• To improve the accuracy and applicability of brain parcellation for diverse neuroimaging data.

Main Methods:

• Propose "pseudo-render-inverse-render" (PRIR) using a direct-inverse rendering framework.
• Render meshes from multiple views, projecting normal vectors to 3-channel images.
• Utilize U-Nets for 2D image segmentation and map results back to 3D vertices.
• Employ Markov Random Fields for postprocessing to ensure smoothness and handle occlusions.

Main Results:

• PRIR is independent of mesh topology, unlike traditional and deep learning methods.
• Achieves state-of-the-art accuracy on topologically correct white matter meshes.
• Demonstrates accurate segmentation for simulated and real surface scans, including noisy data.

Conclusions:

• PRIR offers a robust and versatile solution for cortical mesh parcellation.
• The method effectively handles topological imperfections and captures long-range dependencies.
• PRIR advances neuroimaging analysis by enabling accurate segmentation of challenging datasets.