4465

Histology-Derived pseudo-ADC correlates with measured ADC and extranuclear space in a transgenic model of prostate cancer, identifying contribution of luminal space to measured ADC.
Matthew David Blackledge1,2, Konstantinos Zormpas-Petridis1, Andreas Heindl3, Siver A. Moestue4, Yinyin Yuan3, Dow Mu Koh1,2, David J Collins1,2, Yann Jamin1, Tone F. Bathen4, Martin O Leach1,2, and Deborah K. Hill4

1Division of Radiotherapy and Imaging, The Institute of Cancer Research, London, United Kingdom, 2MRI Unit, The Royal Marsden NHS Foundation Trust, Sutton, United Kingdom, 3Division of Molecular Pathology & Centre for Evolution and Cancer, The Institute of Cancer Research, London, United Kingdom, 4Institute of Circulation and Medical Imaging, NTNU - Norwegian University of Science and Technology, Trondheim, Norway

Synopsis

In this preclinical study we investigate the utility of in-silico simulations of the pseudo-apparent diffusion coefficient (pADC) of water within extra-nuclear regions segmented on large field-of-view haematoxylin, eosin, and saffron (HES) slides from a transgenic mouse model of prostate cancer. We demonstrate that pADC is correlated within in-vivo measurements of the apparent diffusion coefficient (ADC) measured by diffusion-weighted magnetic resonance imaging and may thus be used as a surrogate for exploring the effect of tissue structure on measured ADC values. Furthermore, we demonstrate that ADC is correlated with fractional space occupied by lumen, derived from semi-automatic segmentation of HES slides.

Background

The biological basis of the apparent diffusion coefficient (ADC) as measured by diffusion-weighted imaging is poorly understood in cancer. Previous studies have demonstrated a correlation between ADC and cellular density (number of cells per unit area) measured from whole-slide histological specimens in a preclinical model of prostate cancer1. The prostate is a glandular structure containing luminal fluid-filled spaces; water diffusion occurs more freely in these spaces compared to surrounding cellular regions and the presence of luminal spaces may contribute to higher tissue ADC values2. In this study we demonstrate that in silico simulations of the diffusing trajectories of molecules within haematoxylin, eosin, and saffron (HES) stained slides from whole-prostate can be used to investigate the relationship between the structural properties of the underlying tissue (such as cellularity, luminal space, and extranuclear fraction) and the ADC value measured by MRI. From simulation we derive a histological pseudo-Apparent Diffusion Coefficient (pADC) and quantify the relationship with luminal space, estimated from segmentation of HES samples, and MR-derived ADC.

Materials and Methods

A cohort of nine mice from a transgenic model of prostate cancer (TRAMP) and six healthy control mice (C57BL/6) were imaged on a 7T, preclinical imaging system (Bruker Biospec, Germany). Mice were terminated immediately after image acquisition; upon sacrifice the genito-urinary (GU) tract was excised, formalin-fixed and paraffin embedded for HES staining (slice thickness: 4μm). Histology slides were digitized using a NanoZoomer XR (Hamamatsu, Japan) scanner (40x). Multi-shot EPI, diffusion-weighted imaging was acquired using parameters described in a previous report3 with b-values=0,100,200,400,800 s/mm2, gradient separation Δ=10 ms and gradient duration δ=4 ms (ADC calculated using monoexponential fitting). Digitised histology slides were converted to the LAB colour space and segmented for nuclei using a previously described technique1. For each 500x500 pixel region (0.135x0.135 mm2) in HES slides, the random trajectories of 104 particles were simulated within the extranuclear space of segmented images over a diffusion time that matched MR-experiments (T=Δ-δ/3) to extract a map of pseudo-Apparent Diffusion Coefficient, pADC (Figure 1). From the segmented nuclei, we extracted the fractional area occupied by extranuclear space, ε; similarly, we derived the fractional area occupied by lumen, λ, using a novel HES segmentation approach (Figure 2). An ADC map from a single slice was chosen in each animal that visually best represented the HES cross-section. Corresponding regions of interest (ROIs) were drawn on all maps (Figure 3) and graded into one of four groups according to histological evaluation: Healthy Prostate (C57BL/6), prostatic intraepithelial neoplasia (PIN), well-differentiated (WD) or poorly-differentiated (PD) adenocarcinoma. Using all pixels within ROIs drawn on pADC and ε maps we validated a previously proposed model linking both parameters1,4,5:

$$\varepsilon = \left(\frac{\text{pADC}}{\text{D}_{0}}\right)^{\frac{1}{2\tau}}$$ Eq. 1

where 𝜏 is the ‘tortuosity exponent’ and D0 the diffusion coefficient in the absence of boundaries (data fit using a Markov-Chain Monte Carlo approach to provide parameter confidence). Furthermore, we investigated the correlation between MR-derived ADC and HES-derived pADC and luminal fraction using a least-squares linear fit of mean values within ROIs.

Results

Good visual alignment was generally possible between HES-derived maps and MR-derived ADC (Figure 3). Model fitting of pADC with extranuclear fraction using Eq. 1 (Figure 4) provided an estimate for the extranuclear tortuosity exponent of this mouse population of 0.630 (95% CI: 0.629,0.632); boxplots of data residuals from the derived model reveal no major discrepancies between different animals and/or tissue subtypes (Figure 4). Significant, positive correlations were found when comparing ADC values with pADC (p<0.0005) and luminal space (p<0.005) (Figure 5).

Discussion and Conclusion

Our estimate for the tortuosity exponent in this population is different from a value that indicates linearity between pADC and extranuclear fraction (τ=0.5), providing evidence that simple linear models between both are not appropriate. The linear fit between pADC and ADC reveals a positive bias in the former (positive intercept=0.93), indicating that movement of water molecules, as detected using DWI, is restricted by more than just nuclei (e.g. is sensitive to tissue components such as cell membranes and the extracellular matrix). The strong correlation between pADC and ADC demonstrates that pADC may provide a potent surrogate for identifying the tissue properties that give rise to ADC measurements made in-vivo. By including additional information such as fractional luminal space measured by whole-slide HES technologies, we gain even deeper insight into the structural properties affecting ADC in tissues such as the prostate6. Interrogating tissue structure with these methods we hope to enhance our understanding of longitudinal ADC measurements of different tissue components within multiple cancer types, thus improving our knowledge of tumour heterogeneity and how it may affect treatment.

Acknowledgements

CRUK and EPSRC support to the Cancer Imaging Centre at ICR and RMH in association with MRC and Department of Health C1060/A10334, C1060/A16464 and NHS funding to the NIHR Biomedical Research Centre and the Clinical Research Facility in Imaging. We acknowledge support from the liaison Committee between the Central Norway Regional Health Authority and the Norwegian University of Science and Technology, the Research Council of Norway. MRI was performed at the MR Core Facility, NTNU, animals were housed at the Comparative Medicine Core Facility, NTNU, and histology performed at the Cellular and Molecular Imaging Core Facility, NTNU.

References

1. Hill, D. K., Heindl, A., et al., Proc. 23rd Ann Mtg ISMRM (2016), 0444

2. Chaterjee, A., Watson, G., et al., Radiology (2015), 277(3):751-62

3. Hill, D. K., Kim, E., et al., J. Magn. Reson. Imag. (2016), 43:1207–1217

4. Pfeuffer, J., Dreher, W., et al., Magn. Reson. Imag. (1998), 16(9): 1023-1032

5. Khanafer, K. and Vafai, K. Heat and Mass Transfer (2006), 42(10):939-953

6. Selnæs, K., Vettukattil, R., et al., Frontiers in Oncology (2016), 6:146

Figures

(a) HES images were divided into subregions (500x500 pixels); luminosity channels were derived (b); nuclear boundaries were detected (c, red outlines). Trajectories of 104 diffusing particles were simulated within the extranuclear space over N=1000 time increments for a duration of T=8.67ms (total diffusion time matched to MRI experiments) (d). Each step was sampled from a 2D Gaussian distribution (isotropic variance σ2=2DfT/N with Df = 3x10-3 mm2/s, the diffusion coefficient of free water at 37oC). pADC was estimated using the time-gradient of the mean-square-displacement of particles (e), which represented a single pixel in the final map (f).

(a) k-means clustering of colour channels in the LAB image colour space derived six classes from HES images; a subset of suitable classes for the segmentation of lumen were manually selected, converted to binary and processed using morphological closing and region labeling (b). For classes where the exterior of the lumen was identified, the class-compliment was extracted (c-left); labeled areas above and below a threshold were discarded (d). Identified areas from selected clusters were unified (e) and the fraction of pixels occupied by lumen within 20x20 pixel-regions represented a single pixel in the final luminal fraction, λ-map (f).

The slice-location of MR-derived ADC maps (last column) was chosen such that they visually matched the plane of HES sections (first column). Maps of pADC, extranuclear fraction and luminal fraction were generated from the HES sections (second, third, and fourth column respectively). Corresponding regions of interest were drawn on both modalities: red regions represent disease whilst green represent healthy prostate in C57BL/6 mice and prostatic intraepithelial neoplasia (PIN) in TRAMP mice. Good visual alignment is achieved, and correspondence of image contrast on pADC and in-vivo measured ADC can be observed. WD = well-differentiated adenocarcinoma, PD = poorly-differentiated adenocarcinoma.

A plot of pixel-wise correspondence between histology derived estimates of pseudo-ADC (pADC) against extranuclear area fraction, ε, from all tissue samples. A model linking ε to pADC via the free diffusion coefficient D0 and tortuosity exponent τ fits data from all tissue types and mice in our study (bold-dashed line). This is supported by a box-plot of the residuals for ε (bottom-right) for each mouse and tissue-subtype combination. The estimated value for τ is displayed in the plot, with 95% confidence intervals within parentheses. Model prediction confidence intervals (95%) are shown by the fine dashed line.

Comparisons of MRI-derived ADC with histology-derived pADC (left) and luminal fraction (right): Scatter markers depict mean values within regions of interest, bold dashed lines represent the lines of best fit, grey areas represent the 95% confidence interval for the linear fit, and the fine dashed lines represent the 95% prediction confidence. Parameters from linear fitting are provided in the figure legend, with 95% confidence interval in parentheses. In both cases there is a significant positive correlation between parameters (p < 0.05).

Proc. Intl. Soc. Mag. Reson. Med. 25 (2017)
4465