Benchmarking Geometric Deep Learning for Cortical Segmentation and Neurodevelopmental Phenotype Prediction

Abstract

The emerging field of geometric deep learning extends the application of convolutional neural networks to irregular domains such as graphs, meshes and surfaces. Several recent studies have explored the potential for using these techniques to analyse and segment the cortical surface. However, there has been no comprehensive comparison of these approaches to one another, nor to existing Euclidean methods, to date. This paper benchmarks a collection of geometric and traditional deep learning models on phenotype prediction and segmentation of sphericalised neonatal cortical surface data, from the publicly available Developing Human Connectome Project (dHCP). Tasks include prediction of postmenstrual age at scan, gestational age at birth and segmentation of the cortical surface into anatomical regions defined by the M-CRIB-S atlas. Performance was assessed not only in terms of model precision, but also in terms of network dependence on image registration, and model interpretation via occlusion. Networks were trained both on sphericalised and anatomical cortical meshes. Findings suggest that the utility of geometric deep learning over traditional deep learning is highly task-specific, which has implications for the design of future deep learning models on the cortical surface. The code, and instructions for data access, are available from https://github.com/Abdulah-Fawaz/Benchmarking-Surface-DL.

Surface Vision Transformers: Flexible Attention-Based Modelling of Biomedical Surfaces

“…Subsequently, surfaces were then resampled to a sixth-order icosahedral mesh of 40,962 equally spaced vertices. Experiments were run on both template-aligned data and unregistered (native) data, and train/test/validation splits parallel those used in [32].…”

“…2) Training: The task of GA prediction is arguably more complicated than the PMA task, as it is run on scans acquired around term-equivalent age (37−45 weeks PMA) for both term and preterm neonates, and therefore is highly correlated to PMA at scan. Previous work by [32] have shown the benefit of deconfounding the scan age for the task of birth age prediction, where for all gDL methods an additional 1D convolution was used to incorporate scan age as a confound, before the last fully connected layer used to make the birth age prediction. Here, a deconfounding strategy was employed where the scan age information was incorporated into the patch sequence by adding an extra embedding to all patches in the sequence before the transformer encoder.…”

“…Recent work [32] showed that each architecture generates very different solutions to the same biomedical problem, with variable performance across different tasks.…”

“…Recently, [32] benchmarked a number of geometric deep learning (gDL) methods on cortical phenotype regression and segmentation, and found that these methods typically involve trade-offs between computational complexity, feature expressivity, and rotational equivariance of the models. The objective of this paper is to propose the SiT as a competitive alternative to these frameworks.…”

A Deep-Discrete Learning Framework for Spherical Surface Registration

“…When learning registrations for S3Reg, this generates distortions that must be corrected through averaging warps learned across multiple rotated orientations of the sphere. Recent work showed that contrary to Spherical U-Net, MoNet convolutions (learned from a mixture of Gaussian kernels) could indeed be trained to be rotationally equivariant [9]. Therefore, we develop a new framework for spherical cortical registration based on MoNet, which also takes inspiration from deep-discrete registration frameworks [15,16] designed to learn larger deformations than deep regression frameworks.…”