A biomimetic tumor tissue phantom for validating diffusion‐weighted MRI measurements

Purpose To develop a biomimetic tumor tissue phantom which more closely reflects water diffusion in biological tissue than previously used phantoms, and to evaluate the stability of the phantom and its potential as a tool for validating diffusion‐weighted (DW) MRI measurements. Methods Coaxial‐electrospraying was used to generate micron‐sized hollow polymer spheres, which mimic cells. The bulk structure was immersed in water, providing a DW‐MRI phantom whose apparent diffusion coefficient (ADC) and microstructural properties were evaluated over a period of 10 months. Independent characterization of the phantom's microstructure was performed using scanning electron microscopy (SEM). The repeatability of the construction process was investigated by generating a second phantom, which underwent high resolution synchrotron‐CT as well as SEM and MR scans. Results ADC values were stable (coefficients of variation (CoVs) < 5%), and varied with diffusion time, with average values of 1.44 ± 0.03 µm2/ms (Δ = 12 ms) and 1.20 ± 0.05 µm2/ms (Δ = 45 ms). Microstructural parameters showed greater variability (CoVs up to 13%), with evidence of bias in sphere size estimates. Similar trends were observed in the second phantom. Conclusion A novel biomimetic phantom has been developed and shown to be stable over 10 months. It is envisaged that such phantoms will be used for further investigation of microstructural models relevant to characterizing tumor tissue, and may also find application in evaluating acquisition protocols and comparing DW‐MRI‐derived biomarkers obtained from different scanners at different sites. Magn Reson Med 80:147–158, 2018. © 2017 The Authors Magnetic Resonance in Medicine published by Wiley Periodicals, Inc. on behalf of International Society for Magnetic Resonance in Medicine. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

the MR time points, with sections taken out of the water less than 1 hour after the start of the MR scans; the SEM analysis therefore provided a ground truth for the MR measurements at each time point. Note that a new section of the SEM sample was removed at each time point, so that sections undergoing SEM were only dried once as opposed to undergoing repeated cycles of immersion and drying.

Synchrotron-CT acquisition
Sample preparation for scanning by computed tomography Samples of phantom B were prepared for synchrotron-CT (sCT) scanning in wet and dry conditions. In each case, samples measured approximately 1 mm diameter × 5 mm length. Wet samples had been immersed in deionised water for 1 week prior to scanning and then inserted into a 20 µl pipette tip, filled with deionised water. The tubing was positioned against the walls of the tip to restrict movement during scanning and the tip was then sealed with Parafilm. Dry samples could be scanned without a pipette tip. Both samples could then be mounted on the magnetic stage for scanning.

Scanning conditions
Tomography was performed at the Diamond-Manchester Imaging Beamline I13-2 of Diamond Light Source, UK. A filtered (950 µm C, 2 mm Al) polychromatic beam (5−35 keV) with parallel geometry was used for imaging (1). Acquisition of phase contrast tomography data was performed using a pco.edge 5.5 detector (PCO AG, Germany) coupled to a CdWO4 scintillator (0.75 mm thick) positioned 45 mm away from the sample; final magnifications for acquisition were ×1.25 and ×10 magnification (FOVs 6.7 mm × 5.6 mm and 0.83 mm × 0.70 mm, respectively). This provided a pixel size of 2.6 µm and 0.33 µm, respectively. The sample was rotated 180 • with 0.05 • between each exposure (0.1−0.25 seconds). 3600 projection images were acquired for each sample and reconstructed with flat and dark field correction (2).

sCT analysis
The ×10 magnification wet and dry datasets were used to characterise the sphere wall thickness and sphere volume fraction, respectively, due to the differing contrast in images from the two conditions (see Figure 8 in the main text). All analyses were carried out with Fiji (3,4). Manual measurements of the wall thickness were made on 50 spheres, with the mean of 4 measurements taken as an estimate of the thickness for each sphere. Perpendicular lines through a sphere centre were used to guide the placement of the 4 measurements, which were also guided by overlaying the lines on a 3 × 3 mean filtered image where the boundaries were more distinct; this is illustrated in Figure S1 for an example case. The 50 spheres were chosen by randomly selecting 10 slices throughout the acquired volume, and randomly placing 5 small boxes on each slice; one sphere within each box was then measured, as long as perpendicular lines could be reliably placed such that the four positions on the sphere's outer edge were not obviously merged with other spheres.
The final wall thickness estimate was taken as the mean ± SD of the resulting 50 values.
FIG. S1. Wall thickness measurement for phantom B. Four measurements were made for each sphere, with positions guided by perpendicular lines through the sphere centre, as well as by comparing overlays on original (left) and mean filtered (right) images.
The sphere volume fraction was estimated from area fraction measurements made by segmenting the dry-condition images into 'sphere' and 'non-sphere' regions using the Trainable Weka Segmentation plugin in Fiji (5). For this, one slice was picked at random, and sparse manual labelling of 'sphere' and 'non-sphere' regions was performed after cropping excess background (see Figure   S2). This labelled image was used to train a classifier, which generated a fully segmented image for the chosen slice. This classifier was then applied to 49 other randomly selected slices, resulting in 50 images segmented into 'sphere' and 'non-sphere' regions. While there was a tendency in some cases to misclassify the central area of spheres as background, due to the low signal inten-sities in these regions, this could partially be accounted for by applying morphological hole filling to the segmentations. Although some misclassified regions remained, visual assessment of these segmentations indicated that this method consistently produced higher-quality segmentations than alternative methods such as simple thresholding or watershed approaches. On the 50 segmented slices, sphere area fractions, a f , were obtained from 3 square ROIs, whose in-plane size matched that of an MR voxel (234 µm). The final sphere volume fraction estimate was taken as the mean ± SD of the resulting 150 a f values. Note that this reflects both the volume of the sphere wall and the genuine intra-sphere volume, and as such is expected to be larger than the MR-estimated intracellular volume fraction, which is assumed to be insensitive to the sphere wall.

Original image Segmentation
Manual labelling . Quoted values are the median ± IQR, for all phantom voxels (black) and, where applicable, excluding voxels where at least one parameter value was within 1% of the fit constraints (red). The first and third row form Figure 4 in the main text.