How can spherical CNNs benefit ML-based diffusion MRI parameter estimation?

TL;DR

This paper shows that spherical CNNs improve the estimation of tissue microstructure parameters from diffusion MRI by being robust to sampling schemes and leveraging rotational equivariance, reducing training data needs.

Contribution

The paper demonstrates that spherical CNNs are robust to different sampling schemes and utilize rotational equivariance to improve microstructure parameter estimation from dMRI.

Findings

Spherical CNNs outperform fully-connected networks in dMRI parameter estimation.

Spherical CNNs are robust to different diffusion sampling schemes.

Rotational equivariance in spherical CNNs reduces training data requirements.

Abstract

This paper demonstrates spherical convolutional neural networks (S-CNN) offer distinct advantages over conventional fully-connected networks (FCN) at estimating scalar parameters of tissue microstructure from diffusion MRI (dMRI). Such microstructure parameters are valuable for identifying pathology and quantifying its extent. However, current clinical practice commonly acquires dMRI data consisting of only 6 diffusion weighted images (DWIs), limiting the accuracy and precision of estimated microstructure indices. Machine learning (ML) has been proposed to address this challenge. However, existing ML-based methods are not robust to differing dMRI gradient sampling schemes, nor are they rotation equivariant. Lack of robustness to sampling schemes requires a new network to be trained for each scheme, complicating the analysis of data from multiple sources. A possible consequence of the lack of rotational equivariance is that the training dataset must contain a diverse range of microstucture orientations. Here, we show spherical CNNs represent a compelling alternative that is robust to new sampling schemes as well as offering rotational equivariance. We show the latter can be leveraged to decrease the number of training datapoints required.