[Paper Review] Geometric Convolutional Neural Network for Analyzing Surface-Based Neuroimaging Data

Si-Baek Seong, Chongwon Pae+1 more|arXiv (Cornell University)|Aug 2, 2017

Medical Image Segmentation Techniques4 references3 citations

TL;DR

This paper proposes Geometric Convolutional Neural Network (gCNN), a novel deep learning framework designed to analyze surface-based neuroimaging data on non-regular spherical meshes, such as cortical thickness maps. By leveraging multi-shell mesh structures and geometric convolutions, gCNN achieves significantly higher sex classification accuracy than SVM and standard image-based CNNs using only MRI thickness data, demonstrating its potential for disease classification from MRI and fMRI data.

ABSTRACT

The conventional CNN, widely used for two-dimensional images, however, is not directly applicable to non-regular geometric surface, such as a cortical thickness. We propose Geometric CNN (gCNN) that deals with data representation over a spherical surface and renders pattern recognition in a multi-shell mesh structure. The classification accuracy for sex was significantly higher than that of SVM and image based CNN. It only uses MRI thickness data to classify gender but this method can expand to classify disease from other MRI or fMRI data

Motivation & Objective

• To address the limitation of standard CNNs in handling non-regular geometric surfaces like cortical thickness maps.
• To develop a deep learning framework capable of pattern recognition on spherical, mesh-structured neuroimaging data.
• To improve classification accuracy in neuroimaging tasks using only cortical thickness data.
• To enable extension of the method to disease classification from MRI and fMRI data.

Proposed method

• gCNN employs geometric convolutions tailored for data on spherical surfaces, preserving topological structure.
• The model uses a multi-shell mesh representation to capture hierarchical features across cortical surfaces.
• Geometric convolutions are defined using local neighborhoods on the mesh, enabling spatial invariance and feature learning on irregular domains.
• The network architecture integrates pooling and nonlinear activation functions adapted to spherical geometry.
• Training is performed end-to-end using standard backpropagation on the mesh-based representation.
• The method is evaluated on sex classification using only T1-weighted MRI-derived cortical thickness maps.

Experimental results

Research questions

• RQ1 Can a deep learning model effectively classify neuroimaging data on non-regular, spherical surfaces like cortical thickness maps?
• RQ2 How does gCNN performance compare to traditional methods like SVM and standard 2D CNNs on surface-based neuroimaging data?
• RQ3 Can gCNN achieve high classification accuracy using only cortical thickness data without additional imaging contrasts?
• RQ4 To what extent can gCNN be generalized to classify neurological diseases from MRI or fMRI data?

Key findings

• gCNN achieved significantly higher classification accuracy for sex prediction compared to both SVM and image-based CNNs.
• The method demonstrated superior performance using only cortical thickness data from T1-weighted MRI scans.
• The model's geometric convolutions effectively captured spatial patterns on irregular spherical surfaces.
• gCNN's architecture is extensible to other neuroimaging modalities, including fMRI and structural MRI for disease classification.
• The results suggest that geometric deep learning is a promising approach for analyzing complex surface-based neuroimaging data.
• The framework provides a foundation for developing more accurate, topology-aware models in neuroimaging analysis.