4228

Feasibility of Nipah virus-induced lesion detection using 0.05T MRI

Kunal Aggarwal¹, Yu Cong², Ji Hyun Lee³, Venkatesh Mani², Claudia Calcagno², Michael Ray Holbrook², and Sairam Geethanath¹
¹Icahn School of Medicine at Mt. Sinai, New York, NY, United States, ²Integrated Research Facility at Fort Detrick, National Institute of Allergy and Infectious Diseases, National Institutes of Health, Fort Detrick, MD, United States, ³Department of Radiology and Imaging Sciences, Clinical Center, National Institutes of Health,, Bethesda, MD, United States

Synopsis

Keywords: Infectious Disease, Infectious disease

Motivation: Portable, low-field MRI systems may represent the only viable strategy to monitor neurological manifestations of infectious diseases, such as Nipah virus (NiV), in vivo, and in low-resource settings.

Goal(s): To determine the optimal resolution for lesion detection at 0.05T and compare texture and area markers between 3T and 0.05T

Approach: We acquired 3T MR images of three NiV-exposed NHPs, simulated corresponding low-field images for lesion segmentation.

Results: We determined a minimum resolution requirement of 1.5 x 1.5 x 2 mm³ at 0.05T to manually segment NiV-induced brain lesions on T2w images and that 0.05T texture matches better with 3T data than lesion area.

Impact: The study determined the optimal resolution and textural feature required to monitor NiV-affected patients through prospective imaging at 0.05T to track lesions caused by NiV. It showed that textural was more consistent between the two field strengths than lesion area.

Introduction

High-consequence pathogens, such as Ebola, Lassa, and Nipah viruses (EBOV, LASV, NiV), are associated with acute or long-term neurological manifestations that can be visualized using MRI. However, high-field (HF) MRI scanners (3T or higher) commonly used for brain imaging are extremely difficult to deploy in geographic areas where disease outbreaks naturally occur or in their immediate proximity, for long-term monitoring of human survivors. Portable, very low-field (VLF) MRI systems may represent the only viable strategy to characterize and monitor neurological manifestations of infectious diseases in humans, in vivo, non-invasively, and in low-resource settings. In this work, we explore the feasibility of segmenting focal lesions detected in the brain of NiV-exposed NHPs using 3T MRI, and compare it to corresponding results from VLF-simulated images. We also demonstrate the feasibility of acquiring VLF in vivo data at the simulation-directed resolutions for lesion detection. This study will help develop methods to facilitate VLF MRI deployment in challenging settings to profile neurological sequelae of various infectious diseases.

Methods

Animal model and study design: Three NHPs were exposed to a large-particle aerosol (12 µm) challenge of NiV (Malaysia isolate) at the target dose of approximately 500 plaque-forming units (confirmed by back-titration). Animals underwent brain MRI before and at several time points after virus exposure.
MRI acquisition: Animals were imaged on a 3T Philips Achieva scanner. The acquisition parameters were: TR/TE = 10000/100 ms; field of view (FOV) = 112 x 112 x 60 mm³, with a flip angle (FA) of 90°, resolution of 0.5 x 0.5 x 2 mm³. All NHPs were subjected to imaging twice before the NiV exposure. Subsequent imaging was performed every 3 days after exposure for 2 weeks and then weekly until week 8 after exposure to track lesions. A representative lesion from NHP 3 for three chosen time points is shown in Figure 1.
VLF Simulation: We implemented a VLF simulation pipeline to translate 3T lesion analysis to 0.05T. Figure 2 shows the two steps of the VLF simulation of spatial resolution down-sampling and signal-to-noise ratio (SNR) reduction. VLF noise was then compiled based on cropping 16 voxel-wide corner signal intensities found in ten datasets of 3D Turbo Spin Echo acquired at 0.05T (Multiwave Technologies SA, France) after they had been normalized to the range of 0 to 1 and scaled by a factor of 5 to match VLF SNR, similar to the works on native noise in refs (6-8). This noise (σ_{VLF acquired}) was then scaled and added iteratively to the simulated data until the standard deviation of the simulated noise matched that of the acquired 0.05T data (σ_{VLF Practical)}.
Validation of VLF simulation: We acquired a T2w image of the pro-MRI phantom on a 3T Siemens Skyra with a 0.75 x 0.75 x 5mm³ resolution. The same phantom was acquired on a 0.05T (Multiwave Technologies SA, France) using a 3D Turbo spin echo sequence. We passed the HF data into our LF simulator to mimic the 0.05T data for resolution and SNR. In the resize step of our simulations, we defined a target of 1.5mm x 1.5mm x 5mm in line with the 0.05T data resolution.
Lesion analysis: We used the simulator to investigate the effect of resolution at 0.05T required for lesion detection and characterization. We manually segmented the lesions at 3T and 0.05T simulated data, computed their areas and texture (variance of the gray level cooccurrence matrix)

Results

Figure 3 illustrates the results of the SNR simulation and validation, with the actual (23 a.u.) and simulated SNR (24 a.u.) being similar. The use of B0 mapping (3e) mitigated geometric distortion. The different spatial resolutions at the simulated 0.05T (Figure 4) show that a minimum of 1.5 x 1.5 x 2mm³ is required for manual lesion detection. At this resolution, a graduate student manually segmented 20 out of 21 lesions at the simulated 0.05T compared to 3T (Figure 5a). However, the loss of resolution resulted in increased errors in areas compared to 3T (red arrow, Figure 5b). This difference is reduced by using GLCM variance, as shown by the green arrow in Figure 5c.

Conclusion and future work

We have determined an optimal resolution at 0.05T for lesion detection in this preliminary feasibility study. We used textural analysis as a translational imaging marker from 3T to 0.05T, across 21 lesions found in three NHPs. Future work involves prospectively imaging NiV-infected NHPs with more timepoints.

Acknowledgements

No acknowledgement found.

References

1. Chua, K.B., Bellini, W.J., Rota, P.A., Harcourt, B.H., Tamin, A., Lam, S.K., Ksiazek, T.G., Rollin, P.E., Zaki, S.R., Shieh, W.J. and Goldsmith, C.S., 2000. Nipah virus: a recently emergent deadly paramyxovirus. Science, 288(5470), pp.1432-1435.

2. Sharma, Vikrant, Sulochana Kaushik, Ramesh Kumar, Jaya Parkash Yadav, and Samander Kaushik. "Emerging trends of Nipah virus: A review." Reviews in medical virology 29, no. 1 (2019): e2010.

3. Lim, C.T., 2009. MR imaging in Nipah virus infection. Neurology Asia, 14(1), pp.49-52.

4. Sarji, S.A., Abdullah, B.J.J., Goh, K.J., Tan, C.T. and Wong, K.T., 2000. MR imaging features of Nipah encephalitis. American Journal of Roentgenology, 175(2), pp.437-442.

5. Geethanath, S. and Vaughan Jr, J.T., 2019. Accessible magnetic resonance imaging: a review. Journal of Magnetic Resonance Imaging, 49(7), pp.e65-e77.

6. Geethanath, S., Poojar, P., Ravi, K.S. and Ogbole, G., 2021, May. MRI denoising using native noise. In Proc Intl Soc Mag Reson Med (p. 2405).

7. Qian, E., Poojar, P., Vaughan Jr, J.T., Jin, Z. and Geethanath, S., 2022. Tailored magnetic resonance fingerprinting for simultaneous non‐synthetic and quantitative imaging: A repeatability study. Medical Physics, 49(3), pp.1673-1685.

8. Ravi, K. S., Nandakumar, G., Thomas, N., Lim, M., Qian, E., Jimeno, M. M., ... & Geethanath, S. (2023). Accelerated MRI using intelligent protocolling and subject-specific denoising applied to Alzheimer's disease imaging. Frontiers in Neuroimaging, 2, 1072759.

Figures

Figure 1. The lesion size and texture change over multiple time points after inoculating the NiV in AGM 3. The upper row shows the same MR slice of AGM brain to compare the lesion over time. The red arrow points to the shoulder of the AGM. The bottom row shows a magnification of the lesion's area to see the texture changes over time. The time points (days after inoculation) are indicated in yellow.

Figure 2. Low Field Simulator: First, the resolution of the 3T images was matched by downsampling. Then the ‘target’ SNR at 0.05T was calculated based on the B0 dependent signal level equation. This was followed by adding a noise extracted from the previously acquired 0.05T phantom images. The signal scaling and noise addition step was performed in an iterative manner. With every iteration while calculating the SNR, the phantom's mean value was retained. The noise was iteratively added until the SNR of the 3T image matched that of the 0.05T images within a preset tolerance limit.

Figure 3. Comparison of 3T and 0.05T for a single time point. a) 3T T2w phantom at resolution of 0.75 x0.75 x 5 mm³; b) 0.05T simulated image at 1.5 x 1.5 x 5 mm³; c) 0.05T acquired image; d) 0.05T Off-resonance corrected image; e) off-resonance map acquired at 0.05T. The differences between 0.05T simulated (b) and acquired data (c, d) can be attributed to contrast differences and B0 map quality shown in yellow and black arrows. The yellow text in the images are corresponding SNR values (a.u.).

Figure 4. Effect of resolution at 0.05T simulation for delineating NiV lesions. a) Acquired 3T image of AGM 3 with resolution 0.5 x 0.5 x 2.0 mm³; b-f) Simulated 0.05T image of AGM 3 with resolutions shown in white text. The arrow in yellow shows the lesion, while the arrow in red shows the shoulder of AGM 3. The result in e) with resolution 1.5 x 1.5 x 2.0 mm³ is the optimal trade-off between resolution and 0,05T SNR.

Figure 5. a) Plot for number of lesions detected at 3T and simulated 0.05T for the three NHPs; b) Differences in areas due to the loss of resolution at 0.05T; c) The use of textural feature (variance of the gray level cooccurrence matrix) results in better matched comparisons between the two field strengths

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)

4228

DOI: https://doi.org/10.58530/2024/4228