Image Derived Phenotypes (IDPs) - McIntosh-Lab/tvb-ukbb GitHub Wiki

The original UK Biobank pipeline generates various image-based metrics, or IDPs, for evaluating the characteristics of input images, pipeline processing outputs, and derivative files. These IDPs were intended to be a quantitative measure of the quality of processed subjects but mostly describe structural sub-pipeline processing and outputs. To better capture modalities of interest for connectome-based modelling, we have developed an additional 75 unique IDPs that describe fMRI and dMRI processing as well as connectivity outputs [Supplementary Table 1]. Notable examples include IDPs for assessing the alignment of various modalities to T1 space, the temporal signal-to-noise ratio (tSNR) in resting-state fMRI, and summary statistics for functional and structural connectivity. In conjunction with the original IDPs, these new metrics were developed for the purpose of flagging subjects whose outputs’ quality is poor, either due to acquisition errors, subject anomalies, or pipeline errors and insufficiencies.

We performed a manual QC of 140 (70 female, 70 male) Cam-CAN subjects using our QC reports (described below) to enable assessment of the utility of our newly developed IDPs for quantifying processing errors. The subjects were pseudorandomly selected, balanced for sex and 20 were chosen from each age decile to cover the entire age range of the dataset. Two experienced subject raters (DS, AK) scored the processing intermediates and outputs. These graders gave each subject a score along a 5-point scale for each modality (ranging from excellent [1] to poor [5]) and also gave each subject a pass/fail classification based on the integrity of the TVB inputs as a whole. A fuller description of the QC procedure and example QC report usage for the Cam-CAN data is presented in the Results section.

We used a multivariate statistical approach, partial least squares analysis (Krishnan et al., 2011), to identify a set of latent variables that represent the maximal covariance between the QC ratings and the image-based metrics outputted from the pipeline. First, the covariance between the two sets of variables was computed. Singular value decomposition on this cross-block covariance was then performed to produce latent variables, each containing three elements: (1) a set of weighted “saliences” that describe a pattern of IDPs; (2) a design contrast of QC ratings that express their relation to the saliences, and (3) a scalar singular value that expresses the strength of the covariance. The mutually orthogonal latent variables are extracted in order of magnitude, with the first latent variable explaining the most covariance between the IDPs and QC ratings, the second LV the second most, and so on. We report the relative percentage of total cross-block covariance explained by each latent variable, where the sum of this percentage across all latent variables is 100. The statistical significance of each latent variable was assessed with permutation testing: 1000 permutations shuffled subjects’ QC ratings without replacement while maintaining their IDP assignments. This resulted in 1000 new covariance matrices which were each subjected to singular value decomposition to produce a null distribution of singular values. The reliability with which each IDP expressed the differences across QC ratings was determined with bootstrapping: 500 bootstrap samples were created by resampling subjects with replacement within each rating class. This resulted in 500 new covariance matrices which were, again, subjected to singular value decomposition. The 500 saliences from the bootstrapped dataset were used to build a sampling distribution of the saliences from the original dataset. The bootstrap ratio for a given IDP was calculated by taking the ratio of the salience to its boostrap-estimated standard error. With the assumption that the bootstrap distribution is normal, the bootstrap ratio is akin to a Z-score and corresponding saliences were considered to be reliable if the absolute value of their bootstrap ratio was ≥ 2.

We performed manual QC of 140 Cam-CAN subjects to enable a preliminary assessment of the utility of existing and newly developed IDPs and summary statistics. This assessment was done using a partial least squares analysis of the IDPs with subjects grouped by the rater’s scores. For the functional sub-pipeline, this analysis returned one significant latent variable (Figure 11) showing how IDPs related to head motion, temporal signal-to-noise ratio, the proportion of signal/noise components, and the distribution of functional connectivity values (e.g., centre, range, shape) to be reliable indicators of resting-state fMRI processing quality (p = 0.001, 83.4% cross-block covariance).

_{Figure 11. Partial least squares analysis of functional sub-pipeline IDPs and summary statistics as a function of the functional connectivity QC rating. The analysis returned a contrast (inset) between good (1 & 2) and bad (4 & 5) scores. The most reliable indicators of a good QC rating included high temporal signal-to-noise ratios, low relative displacement, a higher proportion of ICA components classified as signal, and gaussian-like functional connectivity distributions. IDPs and summary statistics with an absolute value bootstrap ratio >2 were considered reliable (see Methods), and are indicated in bold.}

A similar analysis of the diffusion sub-pipeline IDPs resulted in no significant latent variables. This was likely due to a lack of variability in the quality of the diffusion processing and structural connectivity, where nearly all subjects’ (136/140) diffusion sub-pipeline outputs were judged by our raters to be either excellent (1) or very good (2).