Interpretation of the MELD Graph pipeline results

This guideline explains how to interpret the MELD Graph results.

The main outputs of the MELD Graph algorithm are:

  • the MELD Graph PDF report

  • prediction.nii.gz (the cluster predictions for the whole brain)

  • lh.prediction.nii.gz and rh.prediction.nii.gz (the prediction masks for left and right hemispheres)

We recommend that users:

  1. Interpret the MELD Graph PDF report

  2. Visualise the predicted clusters superimposed on the T1 volume. This will enable:

  • Re-review of the T1 /FLAIR at the predicted cluster locations to see if an FCD can now be seen

  • Performing quality control

Here is a video tutorial on how to interpret the MELD Graph results: Interpretation tutorial.

Interpreting the MELD PDF report

The MELD pdf report and .png images of the predicted lesions are saved in the folder: /output/predictions_reports/<subject_id>/reports

The first page of the report provides an overview of all the predicted clusters on an inflated surface. E.g. inflated

If you want to separately view this image, it is named inflatbrain_<subject_id>.png

Each of next pages of the report provides detailed information about 1 predicted cluster.

The following information is provided for each predicted cluster:

  • An image of the cluster on the volumetric T1 image (red) with the 20% most salient voxels (i.e. the voxels most important to the classifier’s decision) in orange

  • The hemisphere the cluster is on

  • The surface area of the cluster (across the cortical surface) in cm2

  • The cortical region in which the cluster mass centre is located

  • The confidence score of the predicted cluster (in %)

  • The integers values used to labelled this cluster and its salient vertices on the NIfTI prediction file.

  • The z-scores of the patient’s cortical features averaged within the cluster. In this example, the most abnormal features are the intrinsic curvature (folding measure) and the sulcal depth.

  • The saliency of each feature to the network - if a feature is brighter pink, that feature was more important to the network. In this example, the intrinsic curvature is most important to the network’s prediction

Viewing predicted clusters

For each predicted cluster, the prediction on the volumetric T1 image (red) with the 20% most salient voxels (i.e. the voxels most important to the classifier’s decision) in orange is displayed. Here is an example:

mri

Each cluster has its own image e.g. mri_<subject_id>c1.png for cluster 1 and mri<subject_id>_c2.png for cluster 2.

Why did the classifier predict this area?

To understand more about the area predicted and why the classifier predicted this area, the report provides the following information:

  • The hemisphere the cluster is on

  • The surface area of the cluster (across the cortical surface) in cm2

  • The cortical region in which the cluster mass centre is located

  • The confidence score of the predicted cluster (in %)

  • The integers values used to labelled this cluster and its salient vertices on the NIfTI prediction file.

  • The z-scores of the patient’s cortical features averaged within the cluster. In this example, the most abnormal features are the intrinsic curvature (folding measure) and the sulcal depth.

  • The saliency of each feature to the network - if a feature is brighter pink, that feature was more important to the network. In this example, the intrinsic curvature is most important to the network’s prediction

We recommend that you carefully review all this information to underd more about the area predicted and why it was predicted. This will help you to decide whether the area identified is likely to be a FCD

For example: saliency

The features that are included in the saliency image are:

  • Grey-white contrast: indicative of blurring at the grey-white matter boundary, lower z-scores indicate more blurring

  • Cortical thickness: higher z-scores indicate thicker cortex, lower z-scores indicate thinner cortex.

  • Sulcal depth: higher z-scores indicate deeper average sulcal depth within the cluster

  • Intrinsic curvature: a measure of cortical deformation that captures folding abnormalities in FCD. Lesions are usually characterised by high z-scores

  • WM FLAIR: FLAIR intensity sampled at 1mm below the grey-white matter boundary. Higher z-scores indicate relative FLAIR hyperintensity, lower z-scores indicate relative FLAIR hypointensity

  • GM FLAIR: FLAIR intensity sampled at 50% of the cortical thickness. Higher z-scores indicate relative FLAIR hyperintensity, lower z-scores indicate relative FLAIR hypointensity

  • Mean curvature: Similar to sulcal depth, this indicates whether a vertex is sulcal or gyral. Its utility is mainly in informing the classifier whether a training vertex is gyral or sulcal. Within FCD lesions, it is usually not characterised by high z-scores or high saliency.

If you only provide a T1 image, the FLAIR features will not be included in the saliency plot.

The information hereabove mentioned about each cluster are summarised into the csv file info_clusters_<subject_id>.csv

Viewing the predicted clusters on the T1 and quality control

After viewing the MELD PDF report, it is then important to visualise the predicted clusters on the T1 to see if they are likely FCDs or not.

The predictions are saved as NIFTI files in the folder: /output/predictions_reports/<subject_id>/predictions

It is important to check that the clusters detected are not due to obvious FreeSurfer reconstruction errors, scan artifacts etc.

Note: Docker does not allow GUI interface, therefore to run the QC you will need to have a stand alone installation of FreeSurfer/FreeView to enable the visualisation.

Open a terminal and cd to where you extracted the release zip.

You will need to first activate FreeSurfer

export FREESURFER_HOME=<freesurfer_installation_directory>
source $FREESURFER_HOME/SetUpFreeSurfer.sh

Then run the command:

python scripts/new_patient_pipeline/new_pt_qc_script_standalone.py -id <subject_id> -meld_data <path_to_meld_data_folder>

Open a terminal and cd to the meld graph folder.

You will need to first activate FreeSurfer

export FREESURFER_HOME=<freesurfer_installation_directory>
source $FREESURFER_HOME/SetUpFreeSurfer.sh

Then run the command:

./meldgraph.sh new_pt_qc_script.py -id <subject_id>

This will open FreeView and load the T1 and FLAIR (where available) volumes as well as the classifier predictions on the left and right hemispheres. It will also load the FreeSurfer pial and white surfaces. It should look like that:

qc_surface

You can scroll through and find the predicted clusters.

qc_surface

Example of a predicted cluster (orange) on the right hemisphere. It is overlaid on a T1 image, with the right hemisphere pial and white surfaces visualised. Red arrows point to the cluster.

Things to check for each predicted cluster:

  1. Are there any artifacts in the T1 or FLAIR data that could have caused the classifier to predict that area?

  2. Check the .pial and .white surfaces at the locations of any predicted clusters. Are they following the grey-white matter boundary and pial surface? If not, you need to try and establish if this is just a reconstruction error or if the error is due to the presence of an FCD. If it is just an error or due to an artifact, exclude this prediction. If it is due to an FCD, be aware that the centroid / extent of the lesion may have been missed due to the reconstruction error and that some of the lesion may be adjacent to the predicted cluster.

Note: the classifier is only able to predict areas within the pial and white surfaces.

Limitations

Limitations to be aware of:

  • If there is a reconstruction error due to an FCD, the classifier will only be able to detect areas within the pial and white surfaces and may miss areas of the lesion that are not correctly segmented by FreeSurfer

  • There will be false positive clusters. You will need to look at the predicted clusters with an experienced radiologist to identify the significance of detected areas

  • The classifier has only been trained on FCD lesions and we do not have data on its ability to detect other pathologies e.g. DNET / ganglioglioma / polymicrogyria. As such, the research tool should only be applied to patients with FCD / suspected FCD

  • Performance of the classifier varies according to MRI field strength, data available (e.g. T1 or T1 and FLAIR) and histopathological subtype. For more details of how the classifier performs in different cohorts, see our paper.

Other commands

For native installation only: You can merge the MELD predictions onto the T1 nifti file using the command below. Note that you will need to have FSL installed on your machine.

./meldgraph.sh merge_predictions_t1.py -id <subject_id> -t1 <path_to_t1_nifti> -pred <path_to_meld_prediction_nifti> -output_dir <where_to_save_output>

The command will create the file predictions_merged_t1.nii.gz which corresponds to the predictions masks merged with T1 in RGB format. It can be viewed on RGB viewer or used to transfert on PACS system.

FAQs

Please see our FAQ page for common questions about the interpretation of outputs