3110

Semi-automatic cloud-based workflow for evaluating the central vein sign for MS diagnosis in a multicenter clinical setting
David Moreno-Dominguez1, Marc Ramos1, Daniel S Reich2, Daniel Ontaneda3, Paulo Rodrigues4, and Pascal Sati2

1Neuroimaging, QMENTA Inc., Barcelona, Spain, 2Translational Neuroradiology Section, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, MD, United States, 3Mellen Center for Multiple Sclerosis Treatment and Research, Cleveland Clinic, Cleveland, OH, United States, 4QMENTA Inc., Barcelona, Spain

Synopsis

The central vein sign (CVS) is novel MRI marker for improving the accuracy and reducing the time to diagnose patients with multiple sclerosis (MS). Recent advancements have introduced the MRI FLAIR* contrast, allowing for CVS to be easily identified. In this work, we developed a semi-automatic cloud-based workflow for evaluating the clinical value of the CVS for MS diagnosis using FLAIR* in a multicenter setting. This novel workflow is a powerful tool that has the potential to significantly accelerate the clinical research imaging studies in MS.

INTRODUCTION


The central vein sign (CVS) has recently been proposed as a novel MRI marker for improving the accuracy and reducing the time to diagnose patients with multiple sclerosis (MS)1. Recent advancements in the form of novel MRI techniques have introduced the FLAIR* contrast, combining submillimeter resolution and improved contrast for lesions and veins, allowing for the central vein sign to be easily identified2,3. In this work, we developed a semi-automatic cloud-based workflow for evaluating the clinical value of the CVS for MS diagnosis using FLAIR* in a multicenter setting.

IMAGE ACQUISITION


Brain images of 8 patients suspected of having MS were acquired on a 3T scanner at the Cleveland Clinic. The protocol included the following sequences: 3D T2-FLAIR (1 mm isotropic, TR = 4.8 s, TE = 352 ms. TI = 1800 ms) and T2*-weighted 3D-EPI (0.65 mm isotropic, TR = 64 ms, TE = 35 ms. EPI factor = 15)3. Anonymized DICOM images were uploaded on a cloud-based commercial platform QMENTA accessible through a web browser4.

IMAGE ANALYSIS WORKFLOW


A custom-built neuroimaging analysis workflow was implemented in the QMENTA cloud platform including the following processing steps: raw DICOM data preprocessing, calculation of FLAIR* images, automatic FLAIR-based lesion segmentation, semi-automatic lesion mask quality check (QC), interactive manual annotation of the CVS and automatic calculation of central vein positive lesion rate at the subject-level (Figure 1).

DICOM data uploaded by the imaging center were automatically tagged based on DICOM tags and sequence naming to detect the appropriate T2*3D-EPI and T2-FLAIR input images within the files of each analyzed subject. Once selected, the files were converted into NIfTI format. FLAIR* images were computed as follows: T2* 3D-EPI images were rigidly registered to the MNI 152 template using ANTs software5, as a means of AC-PC alignment and FOV standardization, while remaining in the native resolution. T2-FLAIR images were then rigidly registered to the aligned 3D-EPI images and interpolated to match the latter’s resolution (0.65 mm isotropic). The aligned images were then multiplied to obtain the FLAIR* images. A schema of the processing pipeline is shown in Figure 2 while native and processed images are shown in Figure 3.

Lesion masks were automatically computed from the aligned T2-FLAIR image using the LST-LPA algorithm6. Each disconnected lesion volume from the mask was given a unique identifier and the labeled lesion mask was displayed as an overlay over the T2-FLAIR image, using a viewer derived from Mindcontrol software7. Additionally, the centroid of each lesion volume was pre-computed and listed as shown in Figure 4. When clicking on one of the listed lesions, the multiplanar T2-FLAIR images were re-centered over the selected lesion. During the QC step, each segmented lesion could be accepted (true positive) or rejected (false positive). In the latter case, the lesion volumes were deleted from the masks once QC was completed.

Using a similar display interface, the curated lesions masks were overlaid over the FLAIR* images. The lesion masks allowed to quickly identify the brain lesions to evaluate for the central vein sign. Using various display functionalities (zoom, pan, lesion mask transparency), the CVS could be manually assessed on the multiplanar FLAIR* images for all detected lesions in each subject. Once the CVS was identified in a lesion, the lesion was listed with its spatial coordinates and marked as central vein positive (or CV+), as illustrated in Figure 5. Once the CVS evaluation was completed, the number of CV+ lesion annotations per subject was automatically calculated and saved in a downloadable datasheet that contains the aggregated information of all subjects in the project, including original lesion count and manually detected false positives/negatives.

DISCUSSION/CONCLUSION


A novel cloud-based neuroimaging workflow has been developed for evaluating the clinical value of the CVS for MS diagnosis in a multicenter setting. In this work, we presented its implementation and initial testing at one clinical imaging site. More sites will soon upload their data for reaching a total cohort size of 100 patients (10 patients per site). By using the cloud-based platform, the workflow only requires an internet browser to upload the raw DICOM images and avoids the need for any kind of software installation, thus increasing its usability in a clinical setting. Furthermore, its flexible design allows different individuals/sites to perform various tasks (image upload, lesion segmentation QC, CVS rating) and easily share the results, as well as using the same parameter configurations, in a secure and traceable way. In conclusion, this novel workflow is a powerful tool that has the potential to significantly accelerate the clinical research imaging studies in MS.

Acknowledgements

The Race to Erase MS foundation and NAIMS cooperative are acknowledged for their financial support.

References

  1. Sati P, Oh J, Constable RT, Evangelou N, Guttmann CR, Henry RG, Klawiter EC, Mainero C, Massacesi L, McFarland H, Nelson F. The central vein sign and its clinical evaluation for the diagnosis of multiple sclerosis: a consensus statement from the North American Imaging in Multiple Sclerosis Cooperative. Nature Reviews Neurology. 2016 Dec;12(12):714.
  2. Sati P, George IC, Shea CD, Gaitán MI, Reich DS. FLAIR*: a combined MR contrast technique for visualizing white matter lesions and parenchymal veins. Radiology. 2012 Dec;265(3):926-32.
  3. Sati P, Patil S, Inati S, Wang WT, Derbyshire JA, Krueger G, Reich DS, Butman JA. Rapid MR susceptibility imaging of the brain using segmented 3D echo-planar imaging (3D EPI) and its clinical applications. Magnetom FLASH. 2017;68:26-32.
  4. Lazovski N, Ramos M, Moreno-Dominguez D, Sato T, Peeters T, Prčkovska V, Rodrigues P. Neuroimaging workflow in the cloud : standardizing research. OHBM 2017: 23rd Annual Meeting of the Organization for Human Brain Mapping. 2017. DOI: 10.13140/RG.2.2.30763.75041
  5. Avants BB, Epstein CL, Grossman M, Gee JC. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Medical image analysis. 2008 Feb 1;12(1):26-41.
  6. Schmidt P. Bayesian inference for structured additive regression models for large-scale problems with applications to medical imaging (Doctoral dissertation, lmu). URN: urn:nbn:de:bvb:19-203731
  7. Keshavan A, Datta E, McDonough IM, Madan CR, Jordan K, Henry RG. Mindcontrol: A web application for brain segmentation quality control. NeuroImage. 2017 Mar 30.

Figures

Fig 1: Schematic of the custom-built workflow. User interaction in the workflow is represented by red circles and arrows. Automatic processing units are represented by blue rectangles and data input/output by green ovals.

Fig 2: Schematic of the processing pipeline to compute a FLAIR* image from raw T2* EPI and T2-FLAIR data.

Fig 3: Representative images from a subject: raw and aligned T2*-weighted 3D-EPI images (top row), raw and registered T2-FLAIR images (bottom row), and the computed FLAIR* (right).

Fig 4: Lesion QC multi-planar viewer showing the lesion label overlay over the T2-FLAIR image (right), and the list of disconnected lesion volume centroids (left).

Fig 5: Central Vein annotation view showing the coordinates of the annotated CV+ lesions showing the central vein sign (left), and the multiplanar FLAIR* images with a semi-transparent lesion mask overlay (right).

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
3110