Central Vein Sign Detection Workflow

The Central Vein Sign Detection Workflow is a semi-automatic workflow for the evaluation of the central vein sign (CVS), an advanced imaging biomarker in multiple sclerosis.

This workflow facilitates manual CVS detection via automated post-processing of raw images to calculate FLAIR* images, which offer high contrast between white matter lesions and cerebral veins, as well as automated white matter lesion segmentation.

The Central Vein Sign Detection Workflow consists of the following steps:

  1. FLAIR* (and FLAIR* post-contrast) image computation. 
  2. Input data quality control via Image Quality Score.
  3. MRI Adjudication.
  4. Deep learning-based automated white matter hyperintense lesion segmentation.
  5. False Positive rejection of wrongly detected white matter lesions.
  6. Manual edition of white matter lesion segmentations.
  7. Manual CVS annotation in pre- and post-contrast FLAIR* images.

FLAIR* Image Computation

In the initial step of the workflow, two direct-from-the-scanner MRI sequences - FLAIR (B) and T2* (C) - are combined to produce a FLAIR* image (A) using a post-processing technique developed by Sati et al. in 2012. A FLAIR* post-contrast image is also derived from the analogous post-contrast sequences (not shown). 

The FLAIR* image proves useful for simultaneous visualization of both lesions and veins, as will be required for subsequent CVS detection.

Image Quality Score

The next step is a manual step requiring user interaction. The Image Quality Score analysis, or IQS, is QMENTA’s in-house input data quality control tool. In the context of this workflow, the user is presented with a grid of six MRI images for visual inspection:

  • T1
  • T2 FLAIR
  • T2*
  • T2* post-contrast
  • FLAIR*
  • FLAIR* post-contrast

Left-click and drag the cursor within an image to change the windowing. Use your mouse wheel to scroll through image slices. All images are co-registered and pre-aligned to MNI space to allow synchronous navigation through slices.

Once each image has been inspected, fill in the drop-down forms on the left-hand side of the viewer by selecting the appropriate image quality score for the image: 0 (Non-diagnostic), 1 (Satisfactory), or 2 (Excellent). You can also add free text comments relating to any specific quality findings in the box ‘IQS Comments’. Click Submit to complete the step.

MRI Adjudication

Three identical manual review steps are completed next by three distinct adjudicators. Adjudicators are presented with T1, T1 post-contrast and T2 FLAIR sequences and must contextually assess the images for the presence of dissemination in space and time.

Select the appropriate response in the drop-downs (Yes/No) and click Submit to complete the task.

WMH Lesion Segmentation

This is an automated step involving deep learning-based segmentation of T2 hyperintense lesions in the FLAIR input image. The step invokes QMENTA’s WMH Lesion Segmentation analysis which is based on a Convolutional Neural Network developed in-house. 

For more information on this analysis, see here.

False Positive Rejection

The next manual step facilitates a rapid quality assessment of the previous automatic segmentation step. Here, users are presented with a list of detected lesion coordinates and can quickly reject incorrect lesions, or ‘false positives’, as well as add new ones (‘false negatives’). 

For more information on this step, see this article, paying particular attention to Point mode. A more user-friendly approach to Painter mode is offered in the next step.

WMH Lesion Editor

This manual step provides the user with an opportunity to refine the white matter lesion mask generated in the previous steps using an intuitive online editor. 

For tips on using this editor, see here.

Manual CVS Annotation

In the final steps of the workflow, CVS-positive lesions are annotated. CVS identification is performed in both pre- and post-contrast FLAIR* images. Users simply scroll through the images and place a coordinate marker wherever the biomarker is visible.

For more information on the coordinate annotation functionality, see this article, noting that neither pre-computed coordinates nor an overlay mask are provided in this instance.

Required Inputs

  • T1 
    • Anatomical 3D image
    • Isotropic resolution recommended
    • Must be labelled as ‘T1’ modality
  • T1 post-contrast
    • Anatomical 3D image
    • Isotropic resolution recommended
    • Must be labelled as ‘T1’ modality and have a ‘post_contrast’ tag
  • T2 FLAIR
    • Anatomical 3D image
    • Isotropic resolution recommended
    • Must be labelled as ‘T2’ modality and have a ‘flair’ tag
  • T2*
    • Anatomical 3D image
    • Isotropic resolution recommended
    • Must be labelled as ‘T2-star’ modality
  • T2* post-contrast
    • Anatomical 3D image
    • Isotropic resolution recommended
    • Must be labelled as ‘T2-star’ modality and have a ‘post_contrast’ tag.

Outputs

  • PDF report detailing white matter lesion segmentation results.
  • Metadata values extracted from manual steps including:
    • IQS scores
    • False positives vs accepted lesions
    • Lesion count / volume
    • CV+ annotations
    • Dissemination in space/time

Typical Execution Time

  • 60 - 90 minutes, depending on manual steps.

References