Validation of quantitative MRI metrics using full slice histology with automatic axon segmentation
Tanguy Duval1, Blanche Perraud1, Manh-Tung Vuong1, Nibardo Lopez Rios1,2, Nikola Stikov1,3, and Julien Cohen-Adad1,4

1Polytechnique Montréal, Montréal, QC, Canada, 2Medical Biophysics Center, Oriente University, Santiago de Cuba, Cuba, 3Montreal Heart Institute, Montréal, QC, Canada, 4Functional Neuroimaging Unit, CRIUGM, Université de Montréal, Montréal, QC, Canada

Synopsis

In this work we propose to validate and compare AxCaliber/ActiveAx/Noddi/MTV in the spinal cord using full slice histology with axon/myelin segmentation. High resolution data (150µm/px) were acquired on an ex vivo spinal cord and compared voxel by voxel with histology. We found that q-space metrics were precise enough to distinguish between various fiber distributions. A correlation coefficient of r=0.62 was found between AxCaliber and histology for axon diameter metric. Also, good agreement were found between the different q-space models and with MTV.

Purpose

Many quantitative MRI methods have been proposed to measure the diameter of axons and their density (e.g. AxCaliber [1], NODDI [2] or ActiveAx [3]), as well as the volume of myelin or macromolecules (qMT, MTV [4]). Most of these methods have been validated independently, but the exact diffusion modeling in white matter is still under debate [5,6]. In order to compare and rank the precision of different methods and models, the community needs a ground truth obtained from histology of a broad range of fiber distributions. In this work, we focus on the variety of diameter distribution and densities present in the spinal tracts. AxCaliber, NODDI, ActiveAx and MTV protocols were acquired on an ex vivo cat spinal cord. The precision of the extracted metrics was assessed using full slice histology with automatic axon segmentation. We also make our data publicly available as a basis for future comparisons (http://www.neuro.polymtl.ca/downloads).

Methods

Tissue preparation. A cervical segment of cat spinal cord (perfused and post-fixed with paraformaldehyde 4%) was extracted. After 24h, two contiguous 1cm pieces were cut. MRI: The first piece was scanned on a Agilent 7T animal scanner equipped with 600 mT/m gradients. The tissue was washed in PBS 5 days at 4°C before scanning and inserted into a small glass tube filled with buffered water. A custom-made solenoid coil was used for transmission and reception (S11 ~ -40dB). One axial slice of spinal cord was acquired, matrix 64x64. Resolution was 0.16x0.16x0.20 mm3. Diffusion. A single shot EPI sequence was used: BW=250kHz, TR=2s. AxCaliber. Figure 1.a. shows the qspace sampling (2D sampling perpendicular to the spinal cord). Diffusion parameters were δ=3/8/8/8 ms, Δ = 7/12/25/40 ms, G = [0 .. 849] mT/m (199 increments), TE minimized (36 - 62ms). The minimal model of white matter (hindered and restricted in parallel cylinders) was used. Fitting parameters were (i) axon diameter index [3] (ii) fraction of restricted water (fr), and (iii) apparent hindered diffusion coefficient. NODDI and ActiveAx. Figure 1.b. shows the qspace sampling. Diffusion parameters were δ=3ms, Δ=30ms, 4 shells acquired with bvalue=40/189/1680/6720 s.mm-2, TE=47ms. 796 directions were acquired in 27min in each protocol. ActiveAx results were obtained using a two compartment model ZeppelinCylinder model and NODDI used a 4 compartment model WatsonSHStickTortIsoVIsoDot_B0. MTV. Macromolecular Tissue Volume (MTV) was measured using the procedure described in [4]. First a T1 map was produced using an Inversion Recovery Fast Spin Echo [7] with 38 inversion times exponentially distributed between 3ms and 2s (hard-inversion pulse, TR=14s, ESP=6.18ms). Spoiled Gradient-Echo images (2 dummy scan, 16 average, TE=2.4ms, TR=24ms) were acquired using flip angles of 2,4,6,10,15,20,25 and 30°. In order to derive Myelin Volume from MTV we used a scaling factor of 1.65 evaluated from data acquired on monkey corpus callosum [8]. Histology. The second piece of spinal cord was stained with osmium 4%, dehydrated, embedded in paraffin, cut in 4µm slices and imaged using an optical 20x whole slice microscope (Hamamatsu NanoZoomer 2.0-HT). Resolution was 230nm/px and allowed us to segment the axons automatically. We release the MATLAB open-source axon segmentation software initialy developed by [9]: http://www.neuro.polymtl.ca/downloads. Validation and Comparison. This procedure is described in Figure 2. Briefly, the axon-segmented image was downsampled by averaging the axon properties on 150x150µm windows and registered to MRI using affine transformation. Correlation (Pearson coefficient) was computed voxel-wise between MRI and histological metrics.

Results

Figure 3 shows the MRI metrics and histology results, whereas Fig. 4 shows the correlation matrix of MRI and histological metrics. We found a high correlation between AxCaliber and histology for axon diameter (r=0.62) and a moderate correlation for the fraction of restricted water (r=0.38). A good correlation for the measurement of the restricted water fraction between AxCaliber and NODDI (r=0.59), ActiveAx (0.86), FA (0.83) and MTV (0.76).

Discussion

In this work, we used for the first time a fully axon-segmented slice of spinal cord and we compared it with different diffusion and myelin imaging quantitative metrics. We showed, through a correlation analysis, that AxCaliber was precise enough to distinguish between various fiber distributions present in the spinal cord white matter. The relatively low correlation for the restricted fraction and Myelin Volume Fraction could be due to the unequal sensitivity of the axon segmentation software throughout the tissue. Better microscopy (using electron microscope or CARS [10]), as well as manual correction using online crowd contribution [11] would overcome this issue.

Acknowledgements

Funded by the Multiple Sclerosis Society of Canada, the Canadian Institute of Health Research [CIHR FDN-143263], the Fonds de Recherche du Québec - Santé [28826], the Fonds de Recherche du Québec - Nature et Technologies [2015-PR-182754], the Natural Sciences and Engineering Research Council of Canada [435897-2013] and the Quebec BioImaging Network.

We also would like to thank:

- Serge Rossignol, Hugo Delivet-Mongrain and all the members of the SensoriMotor Rehabilitation Research Team (SMRRT) for providing us with the cat spinal cord

- The histology unit of University of Montreal's Institute for Research in Immunology and Cancer (IRIC).

- Marvin Brun-Cosme-Bruny for analysing the data

References

[1] Y. Assaf et al., Magnetic resonance in medicine, 2008, 59, 1347.

[2] H. Zhang et al., NeuroImage, 2012, 61, 1000.

[3] D.C. Alexander et al., NeuroImage, 2010, 52, 1374.

[4] A. Mezer et al., Nature medicine, 2013, 19, 1667.

[5] M. Nilsson et al., Magnetic Resonance Materials in Physics, Biology and Medicine, 2013, 26, 345.

[6] U. Ferizi et al., Magnetic resonance in medicine, 2014, 72, 1785.

[7] J.K. Barral et al., Magnetic resonance in medicine, 2010, 64, 1057.

[8] N. Stikov et al., NeuroImage, 2015, 118, 397.[

9] S. Bégin et al., Biomedical Optics Express, 2014, 5, 4145.

[10] T. Duval et al., in Proceedings of the 23th Annual Meeting of ISMRM, 2015.

[11] C.R.G. Guttmann, spinevirtuallab.org

Figures

Figure 1. qspace sampling of the diffusion protocols. a. AxCaliber. b. Noddi and ActiveAx.

Figure 2. Framework for the comparison of MRI quantitative metrics with histology. The optical microscopy (a) was automatically axon-segmented (b). The image was then downsampled (c) by computing the average axonal metrics in 150x150µm2 pixels. Metrics were finally registered on the MRI (d) using affine transform. MRI quantitative metrics (f) were compared with histology (e) voxel by voxel using Pearson correlation coefficient (g).

Figure 3. Correlation matrix comparing histology (green) quantitative diffusion MRI metrics (orange) and quantitative myelin imaging (blue). Note the numbered cases. Box 1 exhibits a high correlation (0.62) between histology and MRI for axon diameter. Box 2 presents a moderate correlation (0.48) between histology and MRI for the restricted fraction. Box 3 highlights the agreement between 2D and 3D sampled q-pace metrics. Box 4 highlights the link between diffusion and myelin imaging.

Figure 4. Quantitative MRI metrics mapping (left 2 columns) and histology mapping (right column). Top row. Axon diameter metrics. Middle row. Fraction of restricted water. Bottom row. FA and myelin volume fraction metrics.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0928