Improving the robustness of neuroimaging through exploitation of variability in processing pipelines

点击翻译按钮获取中文摘要

ABSTRACT Reproducible findings are essential to scientific advancement. Unfortunately, when fields lack consensus standards for methods, or their implementations, reproducibility tends to be more of an ideal than a reality. Such is the case for functional neuroimaging analysis, where there is a sprawling and heterogeneous analytic space from which scientists can select tools, construct processing pipelines, and draw interpretations from their results. Recent demonstrations of disappointing levels of reproducibility for findings across labs, even when using the same datasets, have made the urgent need to overcome analytic heterogeneity clear. Differences in processing steps, parameters, and their software implementation have all been shown to bias results, limiting their comparability with one another. One solution that has emerged in the literature is the adoption of highly prescribed pipelines, such as the fMRIPrep and HCP Pipelines. While successful in restricting variability, the lack of ground truths or consensus processing components and parameters prevents such efforts from being a desirable long-term solution. An alternative strategy, which our team has successfully deployed to achieve robust results in the face of numerical instabilities, is to develop tools that ensemble results across a space of pipeline configurations (i.e., a range of components and parameters). Based on our prior work, we predict that such a strategy would not only improve the robustness of findings, but minimize biases arising from single pipeline selections that compromise the success of biomarker discovery efforts. We address this challenge by proposing a framework for characterizing, summarizing, and minimizing analytic biases in experimental findings. Building on prior work implementing independently developed pipelines (e.g., ABCD-HCP, CCS, fMRIPrep) within a common platform (i.e., the Configurable Pipeline for the Analysis of Connectomes; C-PAC), we will systematically vary their components to generate a broad space of pipelines (n=192). We will quantify the variability in full-brain functional connectivity matrices generated across configurations, and identify both the contribution of individual components (e.g., segmentation, spatial normalization) and the relationships between pipelines (Aim 1). We will construct robust estimates of functional connectivity by sampling the variability observed across pipelines (Aim 2), and improve the generalizability of brain-phenotype relationships through the extension of machine learning ensembling techniques (Aim 3). We will increase the accessibility of our approach by sampling the pipeline configuration space to identify a minimal set of representative pipelines. The strength of these techniques will be demonstrated by identifying generalizable brain-based biomarkers of cognitive and psychiatric wellness using the NIH ABCD Study dataset. This project will lead a shift in neuroimaging towards the capture and inclusion of dominant sources of variability in functional neuroimaging, and in doing so, help to carry functional neuroimaging out of the reproducibility crisis into an era of robustness. Consistent with the values of open science, all contributions will be made publicly and freely available.