MIND the gap: methodological considerations and guidance for structural MRI similarity network analysis with MIND
Structural similarity networks quantify the similarity of structural properties across cortical regions, providing a macroscopic window onto the organisation of cortical architecture. Morphometric inverse divergence (MIND) is a multivariate metric of similarity between cortical areas, based on the Kullback-Leibler (KL) divergence between areal distributions of multiple MRI features or morphometric variables locally measured at voxel or vertex resolution. MIND has demonstrated technical robustness and biological validity and is increasingly widely used as a measure of cortico-cortical similarity in clinical and developmental network neuroscience. Here we provide in-depth methodological background on KL divergence and MIND, highlighting possible sources of bias, critical user decision points in the design of a MIND processing pipeline, and recommendations for technical risk mitigation in using MIND as a metric of cortical similarity. We use simulated data and observational MRI datasets from adults (UK Biobank, N = 500 T1-weighted and diffusion scans) and neonates (Developing Human Connectome Project, N = 752 T2-weighted scans), to show how the estimator of KL divergence implemented in MIND is potentially influenced or biased by five properties of input MRI feature maps: (i) their smoothness; (ii) the proportion of identical values; (iii) analysis in native or common space and the choice of vertex mesh resolution; (iv) parcellation choice; and (v) covariance between input features. We offer principled and practical guidance for investigators wanting to specify and implement the MIND processing pipeline that is best suited to the constraints and opportunities of the MRI data available to them. These recommendations outline which pipeline steps should be used sparingly, such as vertex map smoothing; which should be used with informed caution, such as parcellation choice or vertex mesh resampling; and which could be newly implemented for more robust estimation of MIND, such as the use of principal component analysis to preprocess multivariate MRI features. To support further development of structural MRI similarity network analysis, and wider adoption of robust MIND methods, we also publish the code used to generate the results in this paper as an open resource.