Interpreting the derivatives of MRIQC¶
What to expect when you are expecting (for your images to be good)¶
This far, you probably have successfully run MRIQC on your dataset. At this point, you have two valuable classes of derivatives: a table of Image Quality Metrics (IQMs) and the MRIQC visual reports. Let’s leverage those outcomes and curate our dataset.
Using the IQMs¶
MRIQC extracts a number of IQMs that can be used as features in machine learning frameworks to predict how a human would have rated the quality of each MRI scan (e.g. our paper).
However, and according to our experience, these IQMs suffer from “batch effects” and require the harmonization of the features across different scanning centers and parameters. In other words, with the data we have collected so far we would expect a performance in accuracy of about 78% for the internal MRIQC classifier in a new (unseen) sample of T1-weighed images.
Since your dataset is likely different to those that were used in training the classifier (available here), then you should expect MRIQC to make mistakes if you apply the internal classifier directly on your dataset, without extending the training of it to your data. For the case of functional MRI, MRIQC does not even provide a pre-trained classifier.
In sum, the IQMs are hard to interpret and wildly unknown (in terms of their typical distributions or, at least, some ranges of values and data moments). For these reasons, we recently started the MRIQC Web-API, a resource for scientist to train new automatic classifiers. We hope that crowdsourcing the IQMs will provide insights over the normative values of these metrics, calculate confidence intervals and discover trends and relationships with metadata such as scanning parameters, or the version of MRIQC.
Therefore the IQMs may not be the most useful tool for scientists to curate their datasets (yet). In other words, MRIQC does not avoid (yet) the time consuming visual assessment of all images in a sample. For that reason, MRIQC generates thorough visual reports to minimize the time it takes for an expert to asses an image. The next step using MRIQC is revising all images of your sample individually. Finally, we will propose some “shortcuts” for when the visual inspection of every image is not possible.
Visual inspection of an MRI scan using the reports¶
Assessing T1-weighted images¶
The “T1w - zoomed” plot¶
This is the one report that experts inspect more thoroughly, in search of:
Abnormal physiology to label incidental findings
Traces of head motion: parallel intensity ripples crossing the cortical sheet in anterior and posterior lobes. For example, a subtle presence of these ripples in our example, at the axial cut 102 suggest a moderate bulk-motion of the participant during scanning.
Traces of head motion and eyes movement: revealed by signal leakage from the eyeballs ghosting along the phase-encode direction of the scan.
Extreme INU derived from strong bias field or failures of receiving coils. Our example 50052 shows this problem in the posterior lobe and the cerebellum. The artifact is generally easier to spot in the coronal view.
Thermal noise, spike noise, and poor tissue (GM/WM) contrast. Example 50052 shows a poor contrast, as compared to e.g. participant 50805). A different occurrence of poor tissue contrast (revealed as a blur effect) is shown by participant 50603.
Ferromagnetic, foreign materials may provoke signal loss and distortions (e.g. the second row of this figure).
Wrap-around sometimes may produce signal overlap from anterior to posterior and vice-versa. This effect is oftentimes accompanied by FoV clipping (see this example).
The “T1w - noise enhanced” plot¶
Some artifacts express at very low intensity levels and are only visible in the background for the absence of signal of interest. Therefore, the background can be used as a baseline of the noise and structured artifacts that are probably present across the object of interest in the scan.
Additionally, bulk head motion is typically highlighted in this view. When this artifact is present, it is generally easy to identify a rippling shades as a ghost of the forehead and back of the cranium, in the air immediately surrounding the head.
Finally, some other undesired aspects can only visible with this view. For instance, the rotation of the image information that is typically applied to align the AC-PC plane and the axial plain of the image grid. This is a deterioration factor of the image since it requires a resampling (re-gridding) of the image. It is generally preferable to set an appropriate s/q-form matrix to the Nifti header.
Additional plots generated using the
Extra visualizations will be generated if MRIQC is run with the