To ensure our structural variables of interest are accurate representations of each individuals’ actual morphology, we must apply quality control processes that identify low quality scans and then either remove them from our analyses or perform manual corrections on these scans to improve their quality. Errors during image acquisition produce artefacts in MRI scans, including scanner-related and participant-related artefacts. Scanner-related artefacts arise due to mechanical or hardware faults within the imaging equipment, such as inhomogeneity in the external magnetic field, faulty head coils, and improper imaging parameters (Han et al., 2006). These issues become more apparent among larger samples utilising multiple scanners across different testing sites, as observed effects may be site dependent. Participant-related artefacts arise from inter-individual variations and can potentially confound group analyses. Head motion during image acquisition is the most frequently observed participant-related artefact, producing diffuse noise during image acquisition (Savalia et al., 2017). These commonly appear as rings across the brain, and are correlated with demographics, so they can introduce systematic bias.
As cortical segmentation relies upon accurate tissue contrasts, many of the MRI artefacts discussed earlier interfere with recon-all within FreeSurfer, resulting in segmentation errors. Head motion and other image artefacts introduce anomalies in T1-weighted MRI scans which make it more difficult for FreeSurfer to differentiate between tissue types, causing inaccurate segmentation. For example, head motion causes shading and blurring, reducing image quality and signal clarity in MRI (Godenschweger et al, 2016). Similarly, low image quality and signal noise also add ambiguity to FreeSurfer’s automated algorithms when contrasting signal intensities for each voxel.
Gross errors can appear in brain regions prone to low signal, such as the orbito-frontal and anterior-temporal cortices (Khashper, Chankowsky, & Carpio-O'Donovan, 2013), due to the local magnetic field changes caused by the sinus cavities. Many segmentation errors within these regions often result in the either the total or partial absence of the anterior temporal lobe from the reconstructed cortical surface as FreeSurfer has failed to identify these regions.
Errors within initial stages of the FreeSurfer processing pipeline negatively impact later processing steps. This commonly arises during skull stripping; failure to adequately remove non-brain tissue from the image causes non-brain tissue, such as dura mater, to be included as grey matter within reconstructions of the cortical surface (Kalavathi & Surya Prasath, 2016). Skull stripping errors arise due to dura mater possessing similar intensity characteristics as grey matter within T1-weighted images. Because FreeSurfer discriminates between tissue using signal intensities, dura mater and cerebral membrane are prone to misclassification as grey matter.
Here is an examp
le of segmentation errors where non-brain tissue is incorrectly classified as grey matter within the reconstructed cerebral cortex:
Tradition methods to identify and fix these errors have commonly involved manually inspecting every single MRI scan within a dataset to locate any segmentation errors or noticeable image artefacts. Once identified, these low-quality scans are then either removed from the dataset or receive manual corrections to improve their quality. An example of manual corrections is shown below:
These can be inconsistent, subjective, and time consuming
This approach was developed by the lab and involves MRIQC and manual corrections
- Step 1:
Obtain IQMs from MRIQC.
- Step 2:
Standardise all IQMs via z-scores and perform a Principal Component Analysis (PCA) on these values. Record the lowest number of components that explains at least 80% of the variance in your original data (e.g. the first n components).
- Step 3:
Obtain z-scores for each resulting principal component (that explains 80% of the original variance) and identify outliers using a threshold of +/- 4 SDs from the mean from each component.
- Step 4:
Obtain Euler’s number for all subjects using the mris_euler_number command within FreeSurfer. This measures the topology of the cortical surface within FreeSurfer using the formula 2 -2n, where n is the number of surface defects. This corresponds well to image quality and lower (more negative) values suggest worse quality. Identify subjects that are -4 SDs away from the mean. This should be estimated on raw Freesurfer surfaces that have not been manually corrected.
- Step 5:
Inspect the surface overlays for each subject flagged in either Step 3 or Step 4. Most of these scans will likely be removed from analyses; however, you might find that there have been false positives in these earlier steps and these scans can remain.
- Step 6:
Produce a brief pdf containing a handful of slices for each subject’s MRI scan. This can be obtained either within MRIQC’s subject level output (the html file for each subject) or can be obtained by viewing and extracting the surface overlays from FreeSurfer to assess segmentation quality specifically. Inspect these surface overlays for each subject to determine if there are any additional scans that need to be removed based on visual inspection. This is done to address any false negatives and identify low-quality scans that were not successfully identified during Step 3 and Step 4.
n.b. This pipeline is still being developed and adjusted. Step 6 in particular is quite time consuming and is being reviewed to find a more efficient approach (Steps 1-5 are good though).