Management of prostate cancer is hindered by the absence of an imaging method that can reliably describe the presence, the grade, and the extent of the disease. Normal prostate glandular tissue is comprised of a system of interconnected ducts lined with secretory epithelium and embedded in a matrix of fibromuscular stroma. The most common form of cancer in the prostate, adenocarcinoma, is characterized by the proliferation of secretory epithelial cells altering the normal gland architecture (1). It is possible that the limited sensitivity and specificity of clinical prostate cancer imaging methods is a result of their essentially indirect method of cancer detection—they measure structural or physiological parameters that are associated with cancer, rather than the microscopic tissue architectural features that are used to make a diagnosis of cancer at histopathology.
An imaging method that generates contrast directly from the microscopic tissue structural changes that distinguish cancer from normal and benign tissue would be expected to provide both sensitive and specific cancer detection. Diffusion-weighted water imaging (DWI) is an obvious candidate for this purpose because the free diffusion of water in tissue is known to be constrained by intracellular and extracellular structures and cell walls. DWI can reveal both the scale and orientation of tissue structure, because contrast depends on the net diffusion of water over a specific time period in a specific direction. The two parameters that are most commonly used to describe the rate and spatial freedom of water diffusion are, respectively, the apparent self-diffusion coefficient or diffusivity, and the fractional anisotropy (2).
DWI studies of prostate tissue in vivo have demonstrated a decrease in the measured apparent self-diffusion coefficient in prostate cancer tissue that correlates (P = 0.017) with Gleason grade (3, 4), which is a standard histopathologic description of architectural change and loss of glandular differentiation (1). The observed decrease in diffusivity in cancer relative to normal glandular tissue has been posited to be consistent with both the loss of high apparent self-diffusion coefficient lumenal and ductal spaces (5, 6), and increased cell density (7).
There is both direct and indirect evidence for the presence of distinct microscopic diffusion compartments in prostate tissue. Indirect evidence comes from reports that when a large range of b values are measured in vivo the diffusion-weighted signal shows a multiexponential decay (8, 9). As well as free and restricted diffusion compartments, multiexponential decay has been hypothesized to result from factors including exchange between restricted diffusion compartments, T2 relaxation effects (10), and macromolecule binding (9).
Recently, direct evidence of three distinct diffusion compartments has come from a preliminary MR microimaging study of formalin fixed prostate tissue (11). This study demonstrated highly restricted diffusion in voxels containing the epithelial cell layer, intermediate diffusivity in the stromal matrix, and free diffusion in ducts and acinar lumina. There was a close correlation between structural features visible in DWI and tissue architecture seen on light microscopy of the same tissue.
The study reported here used diffusion microimaging to perform a quantitative analysis of microscopic diffusion compartmentation in prostate tissue. Our experimental hypothesis was that microscopic diffusion compartmentation is a consistent feature of prostate tissue and that changes in compartmentation parallel tissue structural changes associated with cancer.
MATERIALS AND METHODS
All tissue samples were collected from radical prostatectomy specimens with institutional ethics approval and written informed consent from tissue donors. In total eight samples of normal tissue from five organs and two samples of cancer tissue from two organs were collected from five patients. Whole prostates, immersed ∼ 72 h in 10% neutral buffered formalin postsurgery, were sectioned for routine histopathology. Transverse slices (∼ 4-mm-thick) were examined by a specialist urologic pathologist and full thickness tissue samples were obtained from the left and/or right lateral peripheral zone with a 3-mm core punch (sample volume ∼ 28 mm3). This tissue sample volume is similar to the volume of a typical DWI voxel obtained from the prostate in vivo. The selection of regions for sampling was based on visual assessment of the likely tissue type. The stated cancer grade for the cancer samples (Gleason 3+4) is the most likely grade in the sample location based on histopathologic examination of tissue immediately surrounding the core site. The status of normal tissue samples was confirmed by the same method. In this study, we focused on collection of samples of normal glandular tissue, but for comparison, we also analyzed two tumor tissue samples. Cores were placed in vials of neutral buffered formalin and stored 2–10 weeks at room temperature prior to MR imaging.
Tissue cores were transferred from neutral buffered formalin to phosphate-buffered saline (PBS) containing 0.2% v/v gadolinium contrast agent [dimeglumine gadopentetate (0.5 mg/mL); Magnevist, Schering AG, Germany], giving a final gadolinium concentration of 0.16 mM, and stored overnight at room temperature to wash out formaldehyde. Cores were then removed from contrast/PBS and glued (cyanoacrylate “Superglue”) to a plastic strip to constrain the position of the core during imaging. The core and plastic strip were inserted into a 5 mm diameter NMR tube filled with contrast/PBS solution for imaging. The presence of contrast agent reduces the sample T1 to ∼ 500 ms and enables faster imaging without significant T1 weighting.
Imaging was performed at room temperature (22°C) on a Bruker (Germany) AV700 magnetic resonance microimaging system consisting of a 16.4 T vertical bore magnet interfaced to an AVANCE II spectrometer running Paravision 5 and using a 5 mm solenoid RF coil and Micro2.5 gradient set. For diffusion-weighted imaging, a 3D spin echo DTI sequence with the following parameters was used: TR = 500 ms, TE = 18 ms, number of averages = 1, total imaging time = 14 h, field of view = 8 × 4.5 × 4.5 mm3, acquisition matrix = 200 × 112 × 112 (data resolution = 40 × 40 × 40 μm3). Diffusion parameters: δ = 2 ms, Δ = 12 ms, b = 1200 s/mm2, with six noncollinear directions and two b = 0 images.
Diffusion tensor parameters were calculated using the program Diffusion Toolkit (www.trackvis.org; Ruopeng Wang and Van J. Wedeen; TrackVis.org, Martinos Center for Biomedical Imaging, Massachusetts General Hospital). The resulting image data were displayed and analyzed with MIPAV (Version 0.4.4., mipav.cit.nih.gov; Centre for Information Technology, NIH) and Matlab (Mathworks, Natick, MA). Diffusivity was calculated as the mean of the diffusion tensor eigenvalues. Diffusion anisotropy analysis will be discussed in a separate publication. Differences between means were tested with an independent samples t-test (two-tailed, unequal variance). Gaussian curve fitting was performed with the Matlab curve fitting tool (Method: nonlinear least squares; Robust: On; Algorithm: Trust-region).
Region of Interest Selection
Region of interest (ROI) voxel calculations were based on selection of imaging slices that most clearly permitted unequivocal manual selection of a large area of contiguous voxels composed primarily of a single tissue or compartment type (epithelium, stroma, duct, or tumor). The selection was based on the visual similarity of glandular structure seen in the DWIs and conventional light microscopy of fixed and stained prostate tissue (11). Adjacent slices were checked to minimize partial volume effects. A binary mask made from the regions selected in the DWI slice was used to select voxels from the calculated diffusivity images. As the thickness of normal epithelium is variable (from a minimum of ∼ 15 μm), and may include infolding into the ductal/acinar space, selected “epithelial voxels” may have contained significant partial volumes of ductal space and/or stromal tissue.
High Spatial Resolution Measurement of Diffusion Compartmentation
Diffusion compartmentation observed in eight samples of normal glandular tissue and two samples of Gleason pattern 3+4 cancer is illustrated in Fig. 1. There is significant heterogeneity among normal samples in terms of gland density and the number and size of ducts. In one normal tissue sample (No. 5), it was not possible to select a region that could be confidently assumed to be comprised primarily of epithelium-containing voxels (ROI method). The DWIs of cancer samples are characterized by extensive hyperintense areas and absence of visible ductal spaces—consistent with proliferation of low diffusivity epithelial cells and loss of normal glandular structure.
Figure 2 summarizes measurements of mean diffusivity in the manually selected epithelial, stromal, and ductal compartments of normal glandular tissue, and the low diffusivity (presumptively epithelial) regions of the two samples of cancer tissue (ROI-method). In four of the normal samples, a number of ducts displayed low diffusivity relative to other ducts and the surrounding PBS solution. On examination of adjacent DWI slices, these low diffusivity ducts appeared to have no connection with the surface of the tissue sample. There was more intersample variability in the stromal diffusivity than in the epithelial and ductal measurements. The mean and standard deviation of the tissue sample means was 0.54 ± 0.05 μm2/ms for epithelium-containing voxels, 0.91 ± 0.17 for stromal voxels, and 2.20 ± 0.04 for PBS-filled duct voxels close to the sample surface (consistent with the expected diffusivity of water at 22°C (12)). The differences between means were significant for all pairs of compartments (P < 0.001). There was no significant difference between the “epithelial” diffusivity in the cancer samples and the normal samples (P = 0.36).
Histograms of the mean diffusivity of all voxels in “tissue only” subvolumes (mean volume 10.2 ± 2.9 mm3, 160,000 ± 45,000 voxels) of all samples are shown in Fig. 3. The cancer sample histograms are distinctive in being comprised primarily of a very low diffusivity component (or components).
In order to estimate microscopic compartment volumes, the statistical distribution of voxel diffusivity was analyzed according to the examples shown in Fig. 4. The peak at 2.2 μm2/ms is larger than in Fig. 3, because buffer outside the tissue core is included in the volume. The histogram of mean diffusivity in the normal tissue sample can be accurately described (R2 = 0.999) by three Gaussian components of diffusivity 0.46 ± 0.16, 0.85 ± 0.26, and 2.19 ± 0.29 μm2/ms. Based on the histogram, in this particular sample, voxels with diffusivity less than 0.50 μm2/ms were counted as epithelial, voxels with diffusivity greater than 1.60 μm2/ms were counted as “ductal,” and the remainder as “stromal.”
Figure 5 summarizes the results of fitting three Gaussian components to the tissue only diffusivity histograms (Because the high diffusivity component was very small in tissue only subvolumes the diffusivity and σ parameters were constrained to the values obtained from a prior fit to a larger sample + PBS volume.). Only the two lower diffusivity components are shown, presumptively labeled epithelium and stroma. The third component was consistently centered at ∼ 2.2 μm2/ms and was assigned to PBS-filled ducts. Only one component was needed to fit sample 7 in the low to intermediate diffusivity range. The means of the tissue sample means of epithelium and stroma voxels were significantly different (P < 0.001. 0.45 ± 0.08 μm2/ms for epithelium-containing voxels, and 0.83 ± 0.16 μm2/ms for stromal voxels).
The relative partial volumes of epithelium, stroma, and duct (calculated using cutoffs according to the histogram method shown in Fig. 4) in the tissue only subvolumes of the 10 samples are summarized in Fig. 6. The partial volume of ductal space was less than 5% in all samples but differed widely amongst the normal tissue samples (3.0 ± 1.2%). In the cancer samples 9 and 10, the ductal volumes were 0.20 and 0.14%, respectively. The differences between the normal and cancer tissue were significant for partial volume of all three compartments (epithelium P = 0.0091, stroma P = 0.0149, duct P = 0.0002).
This study provides statistical validation of a preliminary report on microscopic water diffusivity compartmentation in prostate tissue (11), and uses this information to estimate the relative compartment volumes of epithelium, stroma, and ductal space in samples of normal and tumor tissue. Together these two studies have demonstrated that, at least in formalin-fixed prostate tissue, DW imaging can provide strong contrast between tissues comprised primarily of epithelial cells (epithelium), fibromuscular stroma, and ductal spaces. Contrast between epithelium, stroma, and ductal space should be particularly advantageous in adenocarcinoma detection and grading, because the relative volumes and the microscopic structural arrangement of these compartments are the basis of the histopathologic diagnosis and grading of the cancer (1).
Imaging at very high spatial resolution, in this case with ∼ 400,000 times smaller voxels than a typical prostate DWI exam in vivo, permits investigation of earlier speculations about the biophysical basis of contrast phenomena observed in vivo. In the following discussion, we have adapted the anisotropy terminology of Shemesh and Cohen (13) and use the terms microscopic diffusivity (μD) to refer to properties measured at high spatial resolution, and ensemble diffusivity (ED) for properties, measured at low and medium spatial resolution, which are the aggregated result of heterogenous μD.
ROI-Based Measures of Diffusivity
High spatial resolution DWIs (40-μm isotropic voxels) of normal prostate tissue show three compartments with distinct contrast. The tissue structure revealed by these images has been shown to match the structure seen on light microscopy of embedded and stained sections of the same tissue (11) and enables the compartments to be unequivocally labeled as epithelial, stromal, and ductal. We thus have the ability to estimate the partial volumes of these three compartments present in a larger volume such as that typical of a DWI acquisition in vivo. There is, however, the following caveat. On light microscopy, the epithelium lining ductal spaces may not only comprise a layer of thickness ∼ 15–20 μm but also display repetitive folding into the ductal space producing an average thickness up to ∼ 50–75 μm. Thus, many 40-μm isotropic voxels, labeled epithelial will also contain unknown partial volumes of stroma and/or ductal space. We note the qualification that when referring to epithelial voxels we mean, strictly speaking, epithelium-containing voxels.
We obtained estimates of the diffusivities of the stromal and ductal compartments, and of epithelial voxels, by visual examination of DW images and manual creation of masks in which all of the voxels appeared to be of a single type. This method gave quite consistent results (Fig. 2) for epithelial voxels from all samples, including cancer, (0.54 ± 0.05 μm2/ms), but there was considerable intersample variability in stromal diffusivity (0.91 ± 0.17 μm2/ms). These diffusivities are higher than those reported for two normal tissue samples in the preliminary study of Bourne et al. (11) (epithelium 0.4 ± 0.1 μm2/ms, stroma 0.7 ± 0.1 μm2/ms), which are similar to the values found in our number 4 sample.
Our results are in good agreement with the 4.7 T intermediate spatial resolution (500 × 500 × 500 μm3) study of formalin-fixed whole prostates by Xu et al. (14). Our epithelial mean, which includes data from the two cancer samples, is in the middle of the range Xu et al. reported for cancer. Our stromal mean is slightly above the maximum reported for “stromal BPH” (benign prostatic hyperplasia) by Xu et al. (14).
Four of the normal tissue samples contained ducts in which the diffusivity was much lower than in the PBS solution in which the tissue sample was immersed. This was particularly the case in isolated ducts and acini, which appeared to have no direct fluid connection to the exterior surface of the tissue core when their course was followed through adjacent DWI slices. Lowered diffusivity in these ducts and acini may be due to high concentrations of macromolecules. Prostatic secretions are reported to contain ∼ 20 g/L protein (15), and in normal tissue, the ducts and acini contain macromolecular secretory products, which often precipitate as corpora amylacea seen at histopathology. The discrepancy between the fitted Gaussian components and the diffusivity histograms in the range around 1.5 μm2/ms (Fig. 4) may be further evidence of restricted diffusion in some ducts. It is possible that in our sample preparation, which did not involve agitation, these products were only washed out of ducts that were closely connected to the tissue sample surface. This raises the possibility that in vivo the diffusion contrast between ducts and cellular compartments is less than in tissue samples washed in a buffer solution ex vivo. If this were the case the sensitivity of DWI in vivo to changes in ductal compartment volume would be reduced and the assumption of free diffusion in ductal spaces (5, 6) is questionable.
Mean Diffusivity Histograms
The ROI-based measurements of compartment diffusivity depended on visual examination of DW images and manual selection of regions that could be confidently assumed to be comprised primarily of a single tissue or cell type. This selection is based on the visual similarity of glandular structure seen in the DW images and conventional light microscopy of fixed and stained prostate tissue (11). Although this compartment selection method provides estimates of the magnitude of typical diffusivity differences between the compartments (or cell-type groups) that are identifiable on light microscopy, it is not visually possible (nor practically feasible) to assign every voxel to a specific compartment. Consequently, it is not possible to use this method to calculate the volume fraction of each compartment that makes up the total sample. For this purpose, a statistical analysis of the whole sample volume (or a representative large part of it) is required.
The histograms of voxel diffusivity from tissue only subvolumes of each sample could be accurately fitted with three Gaussian components centered at diffusivities similar, but not identical, to the average diffusivities of the epithelial, stromal, and ductal compartments measured by manual ROI selection. We restricted the analyses to tissue only subvolumes of the samples in order to be able to compare the overall volume fractions of the ductal compartment in the samples without bias from the PBS-filled space surrounding the tissue samples. The accurate Gaussian fit to the histogram is consistent with the presence of three physically distinct compartments with different mean diffusivities.
The broad diffusivity peaks in our normal tissue histograms, which include both epithelial and stromal voxels, are centered in the range Xu et al. reported for “Benign PZ.” The centers of the main peaks for the two cancer sample histograms are in the center of the range Xu et al. reported for cancer. There is thus good agreement between the range of μD measured at very high spatial resolution and ED measured at lower spatial resolution, and we can hypothesize that the range of ED reflects variations in the partial volumes of epithelial and stromal tissue in large voxels. As in the ROI-based method, the range of epithelial voxel diffusivities estimated by histogram fitting was small (tissue sample 6 excepted), and the range of stromal voxel diffusivities quite broad. This observation suggests considerable intersample variability in stromal tissue structure.
For most samples, the average mean diffusivity of the epithelial compartment was ∼ 0.1 μm2/ms lower in the Gaussian histogram fit than in the ROI-based estimates. There is no obvious explanation for this discrepancy, but it may result from a combination of variable partial volume effects, even in the high spatial resolution voxels, and the subjective process of manual compartment selection in the ROI-based method.
As noted previously (11), on light microscopy, the epithelial cell layer in normal glandular prostate tissue has an average thickness less than the 40 μm dimension of our high-resolution images. Thus, many low diffusivity epithelial voxels will contain significant partial volumes of ductal space and/or stromal tissue. The actual diffusivity of the epithelial layer, which may include both secretory epithelial cells and basal cells, must be somewhat lower than 0.5 μm2/ms.
Although not directly comparable, our results are not inconsistent with the negative correlation between tissue diffusivity and cell density previously reported by Gibbs et al. (7). The cell density method used by Gibbs et al. was based on measurement of the volume fraction of nuclei visible on light microscopy and does not account for cell type, size, or shape. Epithelial cells have similar cross section to fibromuscular stromal cells but are considerably shorter. Thus, an increase in partial volume of epithelial cells at the expense of stromal cells would result in a decrease in ED, as our data predict, and an increase in “cell density” as reported by Gibbs et al. The biophysical basis of very low diffusivity in tissue comprised primarily of epithelial cells requires further investigation.
Estimation of Compartment Volumes
It was suggested by Bourne et al. (11) that diffusion compartmentation in epithelial, stromal, and/or ductal spaces with distinctly different mean diffusivities is the likely explanation for multiexponential diffusion decay observed in vivo when DWI is performed over an extended b-value range. As a first step to investigating this hypothesis, it is appropriate to develop a method for estimation of the partial volumes of each compartment in a typical volume of tissue in which diffusion decay is measured. In principle, it would be possible to use light microscopy of serial-sectioned tissue samples to estimate compartment partial volumes. However, in practice, this approach would be extraordinarily labor intensive and would be compromised by the difficulty of accurate alignment of the light microscopy images with the MR imaging data.
Taking a μD approach to the problem, we selected tissue only subvolumes (10.2 ± 2.9 mm3) of each of our samples (∼ 40% of the sample volume) and used histogram-based measurements of the average mean diffusivity in the epithelial, stromal, and ductal compartments to calculate partial volumes of each compartment from high resolution 3D diffusivity data.
The overlap of the Gaussian curves representing the individual compartments demonstrates the futility of estimating compartment volume fractions by a method involving visual selection of compartments according to DWI contrast. For the purpose of classifying voxels in order to compute compartment partial volumes, we defined the stromal/epithelial “cutoff diffusivity” for each sample as the diffusivity at the centroid of the area formed by the overlap of the low and intermediate diffusivity Gaussian curves (Fig. 4). This method should result in an equal probability of epithelial voxels being misclassified as stromal and vice versa. The same method was applied to define the stromal/ductal cutoff diffusivity (if there was an overlap).
Compartment partial volumes estimated by this method (Fig. 6) varied widely between normal tissue samples. Nevertheless, the mean partial volume of epithelium in normal tissue was significantly less than the mean in cancer tissue. Similarly, the stromal and ductal partial volumes in normal tissue were significantly larger than the mean in cancer tissue. The partial volume of ductal space was less than 5% in all samples but differed widely amongst the normal tissue samples (3.0 ± 1.2%). The two cancer samples were distinguished from the normal tissue samples by their very low ductal volume (0.20 and 0.14%). Low diffusivity in some ducts, as discussed above, may have resulted in an underestimation of the true ductal volume.
Our observations suggest that, despite intersample variability, DWI can detect distinct differences between Gleason pattern 3+4 prostate adenocarcinoma and normal tissue in terms of partial volumes of epithelium, stroma, and ductal space. Thus, diffusion parameters relate directly to important histopathologic features used for diagnosis of adenocarcinoma. It is reasonable to conclude that proliferation of low diffusivity epithelial cells and concomitant loss of intermediate diffusivity stromal tissue is the basis of the correlation observed in vivo between voxel diffusivity and cancer Gleason grade (3, 4).
Our analysis, although based on over 1.6 million voxels and demonstrating heterogeneous μD, can be considered equivalent to the close examination of just 10 typical in vivo voxels. High-resolution imaging and analysis of a larger number of tissue samples of the size we have studied will be required to establish the natural range of μD in normal and pathological prostate tissue. Nevertheless, we have established a μD-based method of compartment volume estimation that can potentially be used to investigate multiexponential ED signal decay.
Notwithstanding significant differences in tissue state (fixed versus unfixed), imaging system (16.4 T ex vivo versus 1.5 T in vivo), and diffusion weighting parameters, our results are partially consistent with multiexponential diffusion decay measured in vivo (9). In vivo, there was an increase in the fraction of the “slow” (lower diffusivity) component of biexponential fit data in cancer tissue relative to normal tissue, consistent with the higher partial volume of low diffusivity epithelial compartment we found in our two cancer samples. There is, however, less agreement between our observation of consistently similar epithelial diffusivity in both normal and cancer samples and the widely differing diffusivities of both the “slow” and “fast” biexponential fit components for normal and cancer tissue in vivo.
Bourne et al. (11) speculated that, given the distinctly different diffusivities of epithelial, stromal, and ductal compartments, triexponential diffusion decay might be observed under ideal experimental conditions. Even if diffusion compartmentation is a significant contributor to multiexponential signal decay it is likely that the apparently very small partial volume of high diffusivity ductal space would result in extreme difficulty in reliably resolving triexponential versus biexponential behavior.
We suggest that, at this early stage, interpretation of diffusion behavior in vivo on the basis of our studies of fixed tissue ex vivo should be extremely cautious. Although the relative diffusivities of different general types of prostate tissue appear robust to fixation (14), it will also be important to more fully characterize the effects of fixation in order to fully exploit the clinical potential of studies of fixed tissue ex vivo.