Functional MRI Preprocessing in Lesioned Brains: Manual Versus Automated Region of Interest Analysis

Abstract

Functional magnetic resonance imaging (fMRI) has significant potential in the study and treatment of neurological disorders and stroke. Region of interest (ROI) analysis in such studies allows for testing of strong a priori clinical hypotheses with improved statistical power. A commonly used automated approach to ROI analysis is to spatially normalize each participant's structural brain image to a template brain image and define ROIs using an atlas. However, in studies of individuals with structural brain lesions, such as stroke, the gold standard approach may be to manually hand-draw ROIs on each participant's non-normalized structural brain image. Automated approaches to ROI analysis are faster and more standardized, yet are susceptible to preprocessing error (e.g., normalization error) that can be greater in lesioned brains. The manual approach to ROI analysis has high demand for time and expertise, but may provide a more accurate estimate of brain response. In this study, commonly used automated and manual approaches to ROI analysis were directly compared by reanalyzing data from a previously published hypothesis-driven cognitive fMRI study, involving individuals with stroke. The ROI evaluated is the pars opercularis of the inferior frontal gyrus. Significant differences were identified in task-related effect size and percent-activated voxels in this ROI between the automated and manual approaches to ROI analysis. Task interactions, however, were consistent across ROI analysis approaches. These findings support the use of automated approaches to ROI analysis in studies of lesioned brains, provided they employ a task interaction design.

Keywords: inferior frontal gyrus; lesion; region of interest analysis; spatial normalization; stroke.

Figure 1

Left inferior frontal gyrus pars opercularis (BA 44) regions of interest (gray) defined for a participant with stroke according to the (A) automated approach and (B) manual approach, and for a control participant according to the (C) automated approach and (D) manual approach. Automated BA 44 was defined from SPM Anatomy toolbox and is overlaid onto the individual participant’s spatially normalized brain image. Manual BA 44 was defined by-hand based on neuroanatomy and is overlaid onto the individual participant’s non-normalized brain image.

Figure 2

Spatial overlap between ROI maps for the left and right inferior frontal gyrus pars opercularis (BA 44), for control participants and participants with stroke. ROI maps defined manually are displayed in hot (color bar indicates 1–12 participants in each group). ROI maps defined using the automated approach are displayed in blue. Spatial overlap between manual and automated maps is indicated in pink. ROIs are overlaid onto the MNI template brain image in neurological orientation.

Figure 3

Effect size for BA 44 differs between manual and automated approaches to ROI analysis. Effect size for control participants and participants with stroke, for BA 44 in the left and right hemisphere, for (A) right hand action observation and (B) left hand action observation, using the automated approach to ROI analysis (dark gray bars) and the manual approach to ROI analysis (light gray bars). Error bars indicate SEM. *p < 0.05, **p < 0.01.

Figure 4

Brain activation associated with right hand action observation in the left hemisphere of a control participant as evaluated by: (A) a commonly used automated approach to ROI analysis, normalized and overlaid onto the MNI brain image; and (B) a manual approach to ROI analysis, overlaid onto the participant’s non-normalized brain image. ROI masks for the left BA 44 are displayed in gray. In this example, the larger automated ROI in (A) captured a larger number of activated voxels than the smaller manually defined ROI in (B). For display, activation maps are shown at T = 1.67–10 corresponding to p < 0.05 uncorrected.

Figure 5

Brain activation associated with right hand action observation in the left hemisphere of a participant with stroke involving the cortex and internal capsule, as evaluated by: (A) a commonly used automated approach to ROI analysis, normalized and overlaid onto the MNI brain image; and (B) a manual approach to ROI analysis, overlaid onto the participant’s non-normalized brain image. ROI masks for left BA 44 are displayed in gray. In this example, the automated ROI in (A) does not contain the intact tissue from left BA 44, whereas the experimenter was able to demarcate the displaced tissue in the manually defined ROI in (B). For display, activation maps are shown at T = 1.67–10 corresponding to p < 0.05 uncorrected.

Figure 6

Brain activation associated with right hand action observation in the left hemisphere of a participant with stroke involving the internal capsule, as evaluated by: (A) a commonly used automated approach to ROI analysis, normalized and overlaid onto the MNI brain image; and (B) a manual approach to ROI analysis, overlaid onto the participant’s non-normalized brain image. ROI masks for left BA 44 are displayed in gray. In this example, peak activation is localized to the left ventral premotor cortex according to the manual approach in (B), whereas due in part to larger ROI volume and spatial smoothing, the activation is localized to the left BA 44 according to the automated approach in (A). For display, activation maps are shown at T = 1.67–10 corresponding to p < 0.05 uncorrected.