Skip to contents

Volume Quality Metrics and Temporal Weighting

Source: R/volume_quality.R
volume_quality.Rd

Functions for computing volume-level quality metrics (DVARS) and converting them to weights for weighted least squares fitting. This implements "soft scrubbing" - downweighting bad volumes rather than hard censoring them.

Conceptual Overview

Traditional fMRI artifact removal uses hard censoring ("scrubbing"), where volumes exceeding a threshold are completely removed. This loses temporal information and can create discontinuities.

Soft scrubbing instead assigns each volume a weight between 0 and 1 based on its quality. High-quality volumes receive full weight; artifacts receive reduced weight. This preserves temporal continuity while still downweighting problematic data.

DVARS as Quality Metric

DVARS (Derivative of VARiance across voxels) measures the root mean square of the temporal derivative across all voxels: $$DVARS_t = \sqrt{\frac{1}{V} \sum_v (Y_{t,v} - Y_{t-1,v})^2}$$

Interpretation:

  • Low DVARS: signal changed smoothly from previous volume (good)

  • High DVARS: signal changed rapidly (possible artifact or motion)

  • Normalized DVARS: values ~1 are typical; >1.5 suggests artifacts

Weighting Methods

Three methods convert DVARS to weights, offering different trade-offs:

inverse_squared

Formula: w = 1 / (1 + dvars^2)

Properties: Smooth, continuous decay. A volume with DVARS=1 gets
weight 0.5. Most conservative choice - provides gentle downweighting
even for moderately elevated DVARS.

Use when: You want smooth, gradual artifact handling without
sharp transitions.

soft_threshold

Formula: Sigmoid decay above threshold

Properties: Volumes below threshold get full weight; above
threshold, weights decay smoothly. Steepness parameter controls
how rapidly weights drop.

Use when: You have a clear idea of what "acceptable" DVARS looks
like (e.g., 1.5x median) and want to preserve good volumes fully.

tukey

Formula: Tukey bisquare (1 - u^2)^2 for |u| <= 1

Properties: Complete downweighting for extreme values. Volumes
beyond 2x threshold get zero weight. Most aggressive choice.

Use when: You have clear artifacts that should be fully excluded,
similar to hard scrubbing but with smooth transitions.

When to Use Volume Weighting vs Robust Fitting

Both approaches handle artifacts, but through different mechanisms:

Volume weighting downweights entire timepoints uniformly across all voxels. Best when artifacts affect the whole brain (head motion, scanner spikes).

Robust fitting (Huber/Tukey bisquare) downweights outlier residuals voxel-by-voxel. Best when artifacts are spatially localized or when temporal structure matters.

Combined approach: Use volume weighting for global artifacts plus robust fitting for residual voxel-level outliers. Set robust = TRUE, volume_weights = TRUE in fmri_lm().

Typical Usage

For use within fmri_lm(), see the convenience parameter volume_weights or the more detailed volume_weights_options.