Microstructure models for diffusion MRI in breast cancer and surrounding stroma: an ex vivo study
Colleen Bailey1, Bernard Siow2, Eleftheria Panagiotaki1, John H Hipwell1, Sarah E Pinder3, Daniel C Alexander1, and David J Hawkes1

1Centre for Medical Image Computing, University College London, London, United Kingdom, 2Centre for Advanced Biomedical Imaging, University College London, London, United Kingdom, 3Breast Research Pathology, King's College London and Guy's Hospital, London, United Kingdom

Synopsis

A variety of one- and two-compartment models were fitted to rich diffusion data sets from ex vivo breast tissue samples containing tumour. Two compartment models with restriction explained the data better than conventional ADC and bi-exponential models, as determined by the Akaike Information Criterion. In four of seven samples, anisotropy was also observed, although parametric maps of the primary eigenvector direction show that regions of coherence are small (~1 mm diameter).

Purpose

To explore breast microstructure in ex vivo tissue using a wide range of diffusion scan parameters. Microstructural models that included restriction and anisotropy were tested, as well as conventional ADC and DTI.

Methods

Seven breast tissue samples (~1 x 2 x 0.5 cm3) containing a portion of tumour were obtained from the tissue biobank with ethics approval. Samples were immersed in formalin within 30 minutes for a minimum of four months. Samples were rehydrated in phosphate-buffered saline for at least 2 weeks, then transferred to Fomblin Perfluorosolv PFS-1 (Solvay Solexis) for scanning.

Scans were performed on a 9.4 T (Varian Inc) using a 33 mm quadrature coil (RAPID Biomedical). Temperature was maintained at 18.5 ± 0.5°C. Diffusion images were acquired using a pulsed gradient spin echo (3.2 x 3.2 cm2 FOV, 0.25 x 0.25 x 0.5 mm3 resolution). There were a total of 42 DWIs (TR=1 s, TEs and numbers of averages shown in Table 1) in three orthogonal directions + 1 b=0 image, and two DTIs (TR=1 s, TE=45 ms, δ=4.5 ms, Δ=30 ms, G=18.7 and 22.6 G/cm) in 42 directions + 6 b=0 images.

Table 2 shows the one- and two-compartment models tested1,2. The intracellular compartment was assumed to be isotropic, but both restricted (Sphere) and free (Ball) diffusion compartments were tested, including several conventional models (ADC, Bi-exponential). The extracellular compartment was assumed to be hindered but not restricted, ie. Gaussian displacement distribution, but could be isotropic (Ball), a tensor (Tensor) or cylindrically symmetric tensor (Zeppelin). The volume fractions of all compartments sum to 1 and all models included the equilibrium signal and T2 as free fit parameters. The radius of the restricted compartment, R, was constrained between 0.1-20 µm.

Each voxel was fitted separately using a maximum likelihood procedure that accounts for local minima and Rician noise1. The Akaike Information Criterion (AIC), $$$AIC=-2\ln(L)+2k$$$, was calculated, where L is the maximum likelihood obtained from the fit and k is the number of parameters. The parameters for a region of interest (ROI) in fibroglandular tissue (FGT) for each sample are shown as median ± interquartile range (IQR).

Results and Discussion

This is the first study examining diffusion models that incorporate intracellular restriction and anisotropy in breast tissue. Models with restriction better explained the data (Figure 1), consistent with clinical studies showing non-Gaussian diffusion at high b values3,4. Parametric maps demonstrated observable differences in the intracellular volume fraction and radius parameters even when ADC was similar (eg. Figs 2; 3a, b arrows).

In four samples (#2-5), models with restriction and anisotropy (ZeppelinSphere, TensorSphere) best explained the data from most voxels; the remaining three samples, which had smaller amounts of FGT, had more voxels where the BallSphere model explained the data better. Maps of the anisotropy (Fig 3c) showed small regions with coherent direction, which may explain why results from coarser-resolution clinical studies of anisotropy are inconsistent (compare 5–7).

The mechanisms contributing to anisotropy are uncertain: a preclinical breast cancer model noted lower fractional anisotropy (FA) in hypoxic vs normoxic regions8, but the difference was small, ~0.03; an ex vivo study suggested FA was higher in the fibrous stroma than in the remainder of the breast gland9, but did not report values because of low reliability. Both of these studies used a simple DTI model. A two-compartment model that includes an isotropic, restricted component to represent the intracellular space may yield a more accurate measure of the anisotropy in the extracellular space with increased sensitivity to the microstructure. Future work will compare MRI measures with histology.

ADC values were slightly lower than those typically observed in vivo, which may reflect changes in water content, protein cross-linking and temperature. However, the microstructural restrictions and orientation present in vivo should be preserved. Tissue degradation would likely remove barriers to diffusion, whereas single compartment models such as ADC and DTI were insufficient to characterize the diffusion signal.

Although the scan protocol presented here is impractical for clinical use, these results may explain why ADC is often sensitive to the presence of cancer, but has low specificity for grade or degree of invasion. This study provides useful information for improving the specificity of clinical breast protocols.

Conclusion

Models with both restriction and anisotropy best characterized the signal in breast tissue samples containing cancer. The findings provide insight on the origins of diffusion signal and may enable better design of models and scan protocols for more specific information in healthy and diseased breast.

Acknowledgements

No acknowledgement found.

References

1. Panagiotaki, E. et al. Compartment models of the diffusion MR signal in brain white matter: a taxonomy and comparison. Neuroimage 59, 2241–54 (2012).

2. Panagiotaki, E. et al. Noninvasive quantification of solid tumor microstructure using VERDICT MRI. Cancer Res. 74, 1902–1912 (2014).

3. Iima, M. et al. Quantitative Non-Gaussian Diffusion and Intravoxel Incoherent Motion Magnetic Resonance Imaging: Differentiation of Malignant and Benign Breast Lesions. Invest. Radiol. 50, 205–211 (2015).

4. Wu, D. et al. Characterization of Breast Tumors Using Diffusion Kurtosis Imaging (DKI). PLoS One 9, e113240 (2014).

5. Eyal, E. et al. Parametric diffusion tensor imaging of the breast. Invest. Radiol. 47, 284–91 (2012).

6. Partridge, S. C. et al. Diffusion tensor MRI: preliminary anisotropy measures and mapping of breast tumors. J. Magn. Reson. Imaging 31, 339–47 (2010).

7. Tsougos, I. et al. The contribution of diffusion tensor imaging and magnetic resonance spectroscopy for the differentiation of breast lesions at 3T. Acta Radiol. 55, 14–23 (2014).

8. Kakkad, S. M. et al. In vivo and ex vivo diffusion tensor imaging parameters follow Collagen 1 fiber distribution in breast cancer xenograft mode. in Proc. Intl. Soc. Mag. Reson. Med. 22 222 (2015).

9. Norddin, N. et al. Microscopic diffusion properties of fixed breast tissue: Preliminary findings. Magn. Reson. Med. Early View (2015).

Figures

Table 1 The number of averages acquired for each set of scan parameters for 42 DWIs (3 orthogonal directions + b=0). δ is the gradient duration, Δ the time between gradients and G the gradient strength.

Table 2 Models tested (fit parameters in parentheses). Ball = isotropic, unrestricted diffusion. Tensor = unrestricted diffusion in three orthogonal directions. Zeppelin = cylindrically symmetric tensor. Sphere = isotropic diffusion restricted by an impermeable membrane at radius R. Total number fit parameters summarized in the last column.

Figure 1. Fits for a representative sample a) Tensor was unable to account for signal at all gradient strengths (arrows) b) ZeppelinSphere c) ΔAIC, the difference between AIC for a given model and minimum AIC from all models. d) Map of the best model (lowest AIC) for each voxel.

Figure 2. Values for selected parameters across all 7 samples (median +/- IQR) for (a) ADC and (b) ZeppelinSphere models. Comparing eg. samples #3, 5 and 6, fits that produced similar ADC demonstrated differences in parameters such as intracellular volume fraction, fI, and cell radius, R.

Figure 3 Parameter maps demonstrating heterogeneity (a and b). Regions of similar ADC (arrows) showed differences in the intracellular volume fraction, fI, and cell radius, R. c) A vector map of the primary eigenvector, λ1, from the ZeppelinSphere fit with FA coloured by direction showed regions of coherent directionality.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0927