Abstract
We propose a general method for combining multiple models to predict tissue microstructure, with an exemplar using in vivo diffusionrelaxation MRI data. The proposed method obviates the need to select a single ’optimum’ structure model for data analysis in heterogeneous tissues where the best model varies according to local environment. We break signal interpretation into a threestage sequence: (1) application of multiple semiphenomenological models to predict the physical properties of tissue water pools contributing to the observed signal; (2) from each Stage1 semiphenomenological model, application of a tissue microstructure model to predict the relative volumes of tissue structure components that make up each water pool; and (3) aggregation of the predictions of tissue structure, with weightings based on model likelihood and fractional volumes of the water pools from Stage1. The multiple model approach is expected to reduce prediction variance in tissue regions where a complex model is overparameterised, and bias where a model is underparameterised. The separation of signal characterisation (Stage1) from biological assignment (Stage2) enables alternative biological interpretations of the observed physical properties of the system, by application of different tissue structure models. The proposed method is exemplified with human prostate diffusionrelaxation MRI data, but has potential application to a wide range of analyses where a single model may not be optimal throughout the sampled domain.
Introduction
Modelling of MRI data seeks to characterise measured signals in terms of simplified physical and/or mathematical expressions that can be used to predict structural and functional tissue characteristics. Such models can be broadly categorised as either phenomenological or structural.
Phenomenological approaches focus on mathematical descriptions of the signal, and may or may not have biophysical interpretations. For example, in diffusionweighted MRI (dMRI), calculation of an apparent diffusion coefficient (ADC) is based on a model that assumes tissue water has a Gaussian displacement probability, and a consequent monoexponential decrease of signal intensity with increasing diffusion weighting (‘bfactor’):
where S and \(S_0\) are the diffusionweighted and unweighted signals respectively. In biological tissue, such water displacement behavior is rare, due to the structural heterogeneity of tissues at both micro and meso scales. In the prostate, the micro scale would roughly encompass cells and subcellular structures, and the meso scale the highly variable glandular and stromal components present within single measurement voxels. The Gleason grading scheme for prostate cancer^{1} is based on ’glandular architecture’—the arrangement and relative volumes of meso scale tissue components.
Phenomenological parameters are often interpreted to reflect tissue structure or function features. The most common example is the assumption that ADC reflects cell density or ’cellularity’. The cellularity interpretation is widespread despite demonstrations that cell type and tissue microstructure, rather than simple cell density, are major determinants of the calculated ADC^{2,3,4}.
The typically nonmonoexponential signal attenuation in tissue measurements can be characterised by a kurtosis parameter (K):
where D is a ’kurtosisadjusted’ diffusion coefficient. K can be roughly interpreted to characterise a degree of tissue structure heterogeneity that causes a deviation from a Gaussian displacement probability. An alternative ’stretched exponential’ formalism^{5} is based on an assumed continuum of water pools, each having Gaussian behavior:
where D is a ’distributed diffusion coefficient’ and \(\alpha\) a ’stretching factor’.
Any water displacement, resulting from either diffusion or incoherent flow, will contribute to dMRI signal attenuation. Flow effects, generally due to blood flow in microvasculature, are most evident in low bfactor measurements. IVIM (Intravoxel Incoherent Motion) analysis^{6} uses a biexponential signal model to separate vascular flow from true diffusion:
where f is the ’perfusion fraction’ and \(D^*\) is a pseudodiffusion coefficient representing flow in randomly oriented capillaries.
Separate from largedisplacement flow effects, dMRI measurements that include intermediate and high bfactors also often show a distinct biexponential behavior, indicating the presence of distinct ’slow’ and ’fast’ water diffusion pools. These were initially assumed to represent intra and extracellular water, however several studies have demonstrated this to be an overly simplified interpretation, and biexponential diffusion signal behavior has been reported inside single cells^{7,8}. Some biexponential T2 relaxation behavior is thought to reflect distinct pools of ’free’ water and water bound to macromolecules^{9}.
Although these models are essentially phenomenological, they may provide some general insight into tissue microstructure features. The IVIM model helps to resolve vasculature from surrounding tissue, and any signal with distinct biexponential decay behavior suggests the presence of two dominant water pools with distinct average water diffusivities, and minimal exchange of water between the pools. Such water pools need not be physically separated in structural compartments. The distinct observed diffusivities simply reflect structural features that determine water displacement probability over the time scale of the dMRI measurement^{10}.
A dMRI study of whole prostate specimens ex vivo (where perfusion effects are absent) used a ’stretched biexponential’ formalism to demonstrate that the commonly reported two major water pools each display distinctly nonGaussian displacement behavior^{11}. In the human prostate, signal modelling based on T2 relaxation behavior (’Lumenal Water Imaging’), without diffusion weighting, also indicates the presence of two major water pools, with the signal from the longer T2 pool correlating strongly with histological measurements of gland lumen volume^{12}. This correlation does not imply that the longT2 lumen space is a separate ’compartment’ from the shortT2 environment. The short T2 is likely due to high protein/macromolecule concentration, but the model does not interpret this as either an intra or extracellular compartment.
In contrast to purely or semiphenomenological descriptions of MRI signal behavior, microstructural models attempt to define the way the measurement technique (scan protocol) and tissue characteristics interact to produce the measured signal, and to thus predict either specific tissue structural features such as cell size, type, density, shape, and orientation, or the more general features that define pathology. Examples of microstructural models for human prostate MRI include, though are not limited to, VERDICT (Vascular, Extracellular, and Restricted Diffusion for Cytometry in Tumors)^{13} and HMMRI (Hybrid Multidimensional MRI)^{14,15}. These two models will be discussed here because they illustrate some important features and limitations of microstructure modelling. Our exemplar of the multimodel method is based on HMMRI data.
The different measurement techniques of VERDICT and HMMRI create constraints on the structural features that the associated models can predict. The VERDICT protocol includes measurements over incremented diffusion times and diffusion weightings, enabling prediction of cell diameter, intra and extracellular volume fractions, and vascular volume. The HMMRI protocol includes multiple echo times and diffusion weightings to predict the relative volume fractions of epithelium, stroma, and lumen space. An extended version of VERDICT (rVERDICT) adds multiple echo times to increase the ability to resolve tissue compartments based on T2 differences^{16}. Here there is some convergence with HMMRI in terms of scan protocols (data acquisition), while maintaining distinct data modelling. These differences also serve to highlight the important distinction between data acquisition and data analysis choices.
Conceptually, the scan protocol defines and limits the information content of the data. Different data models may extract different parts of the embedded information from the data, and one model may do this better than another for a particular application, however no model can extract absent information. Much of the modelling research literature neglects the possible limitations of scan method, and implicitly assumes that optimisation of imaging applications depends only on determining the best data model for a fixed measurement method^{17}.
Both VERDICT and HMMRI models are simplified descriptors of tissue structure. The model parameters are assigned directly to tissue components despite the known heterogeneity and complexity of the tissue, meaning that multiple tissue features, besides the one assigned, may contribute to the observed model parameter variations.
A wide diversity of microstructure models have been applied to different tissues in recent literature. Their relative performance has been compared in terms of accuracy in prediction of disease (eg. presence and grade of cancer) or histological features, and in more abstract theoretical terms using information criteria. A significant finding of studies that have compared models using information criteria is that there is generally no single ’best’ model, and that the preferred model varies from voxel to voxel throughout the tissue, as demonstrated in brain^{18}, prostate^{19}, breast^{20}, and lymph node tissue^{21}. This finding is unsurprising and entirely consistent with the structural heterogeneity of both normal and diseased tissue, yet presents a dilemma for model choice.
The absence of a single best or optimum model, despite the wealth of literature focusing on relative overall performance of specific single models, suggests the need for an approach that can incorporate the predictions of multiple models with weightings related to the appropriateness of each model for the local tissue environment.
Multimodel averaging methods are used extensively in a wide range of fields, from econometrics to ecology and weather forecasting, yet seem rare in medical imaging^{22,23}. Brix et al.^{24} describe the application of a model averaging technique to the analysis of simulated dynamic contrast MRI data, although not with the specific aim of addressing tissue heterogeneity.
Applied to tissue microstructure prediction based on MRI data, the multimodel approach would appear limited to models that share some specific parameter (eg. intracellular volume fraction or cellularity). In this case modelaveraging could be used to predict a consensus value of the shared parameter (as in Brix et al.^{24}).
In this paper we present a method for multimodel analysis, exemplified using in vivo prostate HMMRI data. By separating the signal description model parameters from specific tissue structure assignment, the method relaxes the conventional direct connection between signal model parameters and tissue structure features  enabling a consensus of diverse signal and structure models, and avoiding the need to select a single ’optimum’ model.
Methods
The method we illustrate here combines the predictions of multiple models that do not share parameters directly representing tissue properties. We separate the conventional ’direct’ microstructure imaging approach into distinct phenomenological/physical and structural/biological modelling steps (Fig. 1). The generic features of the approach are emphasised in italic.

1.
Multimodel semiphenomenological signal analysis that predicts the physical properties of the predominant water pools. For each model, Akaike Information Criterion (AIC) is used to estimate model likelihood. Water pools are not attributed to specific microstructural features. In the example we predict ADC, T2, and fractional volumes (\(f_i\)) of each water pool in prostate tissue.

2.
Application of a single biological microstructure model to predict the relative volumes (\(f_i\)) of tissue components contributing to the individual water pool physical properties predicted at Step 1. In the example, we predict epithelium, lumen, and stroma (ESL) volumes, as for HMMRI. As an alternative example, we predict ’cellularity’.

3.
Microstructure model predictions from Step 2 are combined, with weighting according to model likelihood and water pool fractional volume, to estimate the relative volumes of tissue structural components making up each voxel.
Using previously acquired in vivo prostate HMMRI data, we fit five progressively more complex models that describe signals arising from one, two, or three nonmixing water pools with Gaussian diffusion characteristics and a homogeneous T2 value throughout the pool. The simplest, singlepool, model equates to the conventional ADC model with T2 relaxation. Models 2 and 3 each describe two water pools. In Model 2 both pools have the same T2, while distinct T2 values are allowed in Model 3. These ’biexponential’ models are distinct from the IVIM (intravascular incoherent motion) model^{6} in accommodating T2 relaxation, however, in the example data analysed here, the sampled bvalues (0, 150, 750, and 1500 s/mm^{2}) do not permit accurate assessment of an incoherent perfusion component. Models 4 and 5 describe three water pools with distinct diffusivities. In Model 4 two of these pools share an identical T2 value. Model 5 is equivalent to the mathematical form of the ’standard’ HMMRI model^{14}, but with a broader range of possible diffusivity and T2 values for each water pool (see Supplementary Table S1 online), and no direct assignment of water pools to tissue structure features.
In our exemplar the five models are nested, with the higher order models being less constrained versions of the simpler models. While each of the phenomenological models has a plausible tissue microstructure interpretation (Table 1), this interpretation is not essential, and is not used in this demonstration.
We note that the use of nested models is not a requirement of the method. The only constraint, in our exemplar, is that each model should predict a T2 and D for each water pool.
The conventional HMMRI approach directly and specifically assigns the ESL structural compartments to the three components of the triexponential model. This approach precludes a similar microstructural assignment of the water pools of the mono and biexponential models which may be more appropriate descriptors of the tissue structure in some voxels. To avoid direct assignment, we apply a structural model in the second stage of data analysis to predict the fractional ESL volumes contributing to the predicted water pool D and T2. The microstructural model assumes that the ESL compartments are internally homogeneous but, throughout the prostate tissue, have a Gaussian distribution of D and T2 values. The chosen mean and sigma values for the distributions were based on literature reports and, where available, measurements from defined tissue regions validated by histology (see Supplementary Material online).
By applying the structure modelling to the detected water pools, rather than to the total voxel signal, our method is sensitive to signal variations that would be masked by nonlinear multiexponential decay, while incorporation of mono and biexponential models reduces parameter variance in voxels where the higher order models are overparameterised.
More method details are available in Supplementary Material.
Image acquisition and preprocessing
In vivo prostate HMMRI imaging was performed as described previously^{15}. Briefly, men with biopsy proven prostate cancer were scanned in a 3T system with cardiac phasedarray and endorectal coils. Diffusion weighted images were acquired with all combinations of echo times \(TE =\) 57, 70, 150, and 200 ms, and \(b =\) 0, 150, 750, and 1500 s/mm^{2}, giving a 4 \(\times\) 4 array of diffusionrelaxation measurements for each voxel. The image planes were oriented perpendicular to the rectal wall, and approximately transaxial to the prostatic urethra, as guided by sagittal T2 images, to approximately align MRI planes with whole mount histology slices. Diffusion encoding gradients (\(\Delta \backslash \delta\) = 21/35 ms) were applied in three orthogonal directions with the resultant images averaged. The raw image data used in the current study were also used in previously described work^{25}.
Magnitude mode images were coregistered to reduce motion effects (see Supplementary Fig. S1 online), and Rician noise floor bias was corrected voxelwise according to the method of Gudbjartsson and Patz^{26}.
Step 1. Phenomenological model fitting
Models 1–5 (Table 1) were fitted progressively, starting with the simplest model, with the parameter estimates of the lower order models used to define the parameter limits and starting points for the next higher order model(s) (see Supplementary Fig. S2 online). Nonlinear least squares model fitting was performed in Matlab.
For each voxel we calculated the corrected Akaike information criterion^{27}:
where SSE is the sum of squared residuals, N is the number of measurement data points (16 in our example), and p is the number of estimated model parameters (2–8 in our example). AICc is used in Step 3 to assign a model weight.
Step 2. ESL microstructure model fitting
We assume that, across the prostate and the patient population, for any water pool D and T2 combination the contributions of epithelium, stroma, and lumen (ESL) tissues each have a Gaussian probability distribution with mean and sigma as specified in Table 2. For each water pool T2 and D value predicted by the phenomenological models, we calculate a Zscore (\(Z_E\), \(Z_S\), \(Z_L\); range 0–1) of the Gaussian distribution for the E, S, and L components. The predicted fractional volume (x) of the tissue component contributing to the water pool is then calculated as:
This step predicts the tissue structure (fractional volumes of the ESL compartments) giving rise to the physical properties of each water pool (Fig. 1, ’Predicted tissue composition for each water pool’).
Step 3. Weighted model averaging
The AICc values from the five models (Step 1) were used to calculate model weights (\(W_m\) = normalised likelihood) according to^{28} (p. 447):
where \(\Delta _m\) is the AICc difference between model m and the model with lowest AICc value (\(AICc_m  AICc_{min}\)).
The weighted multimodel ESL prediction for voxel structure was calculated based on the fractional volume of ESL components in each water pool and the associated model weight:
where \(X \in \{E,S,L\}\) is the modelaveraged fractional volume of a given ESL component, m = model number (1, 2, ..., 5), n = number of water pools in model m, \(f_m^p\) = fractional volume of water pool p in model m, and \(x_m^p\) = fractional volume of ESL component x in pool p of model m.
Cancer probability prediction
As a demonstration of the possible application of a pathology model to the predicted ESL composition of the prostate tissue, we calculated a hypothetical cancer probability (\(P_{Ca}\)) for each voxel based on the clinical observation that prostatic adenocarcinoma is characterised by proliferation of epithelial cells and occlusion of lumen space^{29}:
where E and L are the predicted fractional volumes of epithelium and lumen. Stroma volume is omitted as it correlates poorly with presence and grade of prostate cancer^{2}, and is redundant given the epithelium and lumen fractional volumes (\(E+S+L = 1\)). This model is a simple extension of the binary cancer prediction model of conventional HMMRI^{14}.
Alternative ‘cellularity’ model
To illustrate the ease of implementing an alternative microstructure model, we replaced the ESL model with a hypothetical simple cell density model. This model assumes intra and extracellular water having a Gaussian distribution of T2 and D values with mean and sigma as shown in Table 3. In this example ’cellularity’ is presented as intracellular volume fraction, and has a range of 0–1.
Histology data
For the exemplar data presented, there was no precise alignment of histology sections with MRI planes. Radical prostatectomy specimens were sliced in 3 mm thick sections transverse to the posterior surface of the prostate. Thus the histology images lie approximately parallel to the the MRI data, although not necessarily coplanar. We note that rigorous histological validation of MRI studies is an ongoing challenge^{30}, beyond the scope of the general method proposal presented here. In the figures presented here, we show the histology image which has the best overall anatomical match to the MRI features.
Results
Figure 2 shows predicted water pool T2 and D values from the five phenomenological models across a transverse slice of a human prostate that includes a large high grade tumour. The pixel colour represents T2 and D value, while the pixel brightness is weighted according to water pool fractional volume. This weighting emphasises the parameter values of voxels where signal to noise ratio (SNR) for the associated water pool is high and the effect of noise on parameter variance is low. When the fractional volume of a water pool is low the associated SNR will be low, leading to unreliable estimates of T2 and D. These voxels are darkened in the maps. The corresponding unweighted D and T2 maps are provided together with fractional volume maps in the Supplementary Material.
Figures 3 and 4 provide similar data for two further patients.
The T2 maps indicate that for all models the large majority of voxels show water pool T2 estimates that are biopyhsically plausible and lie inside the allowed model fitting range of 10–1000 ms. The twopool models show distinct differences in T2 values between the two pools. In comparison, the threepool models show relatively minor differences between pools 1 and 2, with pool 3 having similar T2 to pool 2 of the twopool models.
The twopool models show distinct diffusivity differences between the pools. Diffusivity values are mostly biophysically plausible, although in Models 2 and 3 there are significant areas with D > 3.1 \(\upmu \hbox {m}^{2}\)/ms (the free diffusion maximum at body temperature). These high D regions, which, based on prostate anatomy, are unlikely to represent true perfusion effects, are much reduced in Models 4 and 5.
All models show clear differences between the large high grade (Gleason \(4+5\)) tumour in the right peripheral zone (PZ) and the noncancerous regions of the prostate. However, the two and threepool model parameter values show much more variation within the benign tissue than the single pool ‘ADC’ model. The generally high T2 values in the PZ are consistent with conventional T2weighted imaging, where benign PZ tissue is typically hyperintense relative to the central prostate. As there was no precise alignment of histology sections and MRI planes^{30}, a close match between the location and boundaries of the tumour in the histology and MR images is not expected.
Figure 5 shows maps of model likelihood, and corresponding predictions of the partial volumes of epithelium, stroma, and lumen space predicted by application of the ESL microstructural model to each of the phenomenological model water pool properties. For the multipool models, the ESL predictions from each pool in the model are averaged, with weighting according to pool fractional volume. In addition, the modelaveraged ESL map is provided with weighting based on individual model likelihoods. The singlemodel ESL maps illustrate how selection of a single model would influence the structure prediction, and the degree to which the prediction from weighted model averaging differs from the single model predictions.
To demonstrate the ability to apply alternative structural models to the multimodel water pool predictions, results from the alternative ’cellularity’ structural model are shown in the third row of Fig. 5. In this model a cellularity of 1 indicates that all water is intracellular, and a cellularity of 0 indicates that all water is extracellular (zero cell density).
The likelihood maps indicate that Model 3 provides the best signal description in a little more than half the voxels, with the majority of remaining voxels best described by the simpler twopool Model 2 or the single pool Model 1. The threepool models have very low likelihoods in all but a few voxels—a result that is unsurprising given the small number of measurements relative to the number of model parameters, and the relatively low SNR of the in vivo measurements.
Despite their low likelihoods, the threepool models predict very similar ESL values to the twopool models—suggesting that the multistep structure modelling approach may compensate overparameterisation.
The weighted multimodel ESL map shows small differences from each of the individual model ESL maps, and patterns that are consistent with the approximately aligned histology section. Notably, highly cystic (acellular) regions of the PZ and central zone roughly correspond to high lumen volume regions (dark blue) in the ESL maps. Model 1 predicts very high stromal density throughout the nontumourous prostate—a result not supported by the histology section. In contrast, the two and three pool models, and the weighted combination, predict a heterogeneous mixture of lumen and stroma, in good agreement with the histology and normal prostate anatomy.
Figure 5 also presents a possible pathological interpretation of the multimodel ESL prediction by calculating a hypothetical cancer probability based on fractional volumes of epithelium and lumen space (Eq. 5). Notwithstanding the imperfect alignment of histology section with MRI slice, the probability map shows a qualitatively good agreement with the histologically defined cancer extent.
The cellularity model correctly predicts high cell density in the tumour, but in contrast to the ESL model, fails to reflect normal anatomical heterogeneity in the peripheral and central regions of the prostate.
Results from the two other patients are provided in Figs. 6 and 7 . The broadly consistent and anatomically plausible ESL predictions across the three patients suggest that the assumptions of T2 and D statistical distributions used for the structure model (Table 2) are reasonable.
Figure 8 provides a qualitative comparison of ESL structure predictions from the multimodel method and the ’conventional’ threecompartment HMMRI model^{15} in three patients. As the alignment of histology sections and in vivo MRI slice planes is imperfect, close correspondence of MRIpredicted structure and histology is not expected. However, based on general prostate anatomical features, there appear to be some advantages of the multimodel approach. In Prostate 1 (i.e., from patient 1), the HMMRI method appears to overestimate the fractional volume of epithelium (bright red) in the tumour, and the fractional volume of lumen space (dark blue) in the posterior peripheral zone. HMMRI also, implausibly, suggests high epithelium volume at the left and anterior edges of the prostate. In Prostates 2 and 3 the multimodel prediction appears to show the dense part of the left side tumour more distinctly than the HMMRI prediction.
Discussion
We have proposed a method for predicting tissue microstructure from MRI data based on combining the predictions of multiple models, with weighting according to calculated model likelihood. Conceptually, our method avoids direct assignment of tissue structure features to signal characteristics. The multiple phenomenological models account for both meso and micro scale heterogeneity of the signal source (usually one voxel) in terms of pools of water that may, or may not, have distinct physical properties. We then apply a single tissue structure model to predict the microstructural features that give rise to the signal behavior of each detected water pool, and aggregate these pool predictions to summarise the structure of the signal source (voxel).
Our approach differs significantly from Brix et al’s multimodel analysis of simulated dynamic contrast enhanced MRI data^{24}, in that we do not directly assign tissue properties to the model parameters, and that we apply the likelihoodbased model weightings, not to the signal models, but to the tissue structure predictions made subsequent to the initial semiphenomenological signal modelling. Below we discuss some of the underlying assumptions and potential refinements and modifications of our approach.
Choice of the five phenomenological models was based on what we considered likely biophysical properties of water pools in prostate tissue. Models with constrained/shared diffusivity and free T2 were not tested as we assumed that while many equally proteindense tissue structures may exist, giving similar T2 but varying D, the alternative of equal diffusion restriction/hindrance but varying protein density is unlikely in animal tissue. Either way, we emphasise that these five models are presented simply to illustrate the general multimodel approach, and the models are not specifically nor exclusively appropriate to the prostate tissue from which the exemplar data were acquired.
Our phenomenological ’water pools’ approach accounts for mesoscopic tissue heterogeneity. Numerous experimental works have demonstrated that multiple signal components are often simplistically overinterpreted as representing distinct microstructural compartments. However, these studies are mostly based on mesoscopically homogeneous tissues and cell suspensions. Many tissues are highly heterogeneous at the meso scale, such that in any single voxel there may be large regions of very different structure. Fig. 9 illustrates such meso scale structure heterogeneity in a 2 \(\times\) 2 mm section of prostate tissue that could represent a crosssection of a typical MRI voxel. This section, as a 3D volume, could plausibly give rise to three distinct water pool signals: (1) free water in the large cystic spaces; (2) highly hindered and restricted water in the dense stroma; and (3) water in the complex of small glands. At the microscopic scale, both the stroma and glandular spaces would be internally heterogeneous, and very different from each other. A threecomponent signal model might directly and erroneously interpret these three water pools as ‘intracellular’, ‘interstitial’, and ‘lumenal’, on the basis of distinct average diffusivities, when the first two pools actually contain both intracellular and interstitial water, and the second water pool contains lumen spaces. This example illustrates how, depending on mesoscopic tissue structure and voxel size, distinct water pools (signals) may be observed and yet be inaccurately characterised by an applied microstructure model.
While we did not explicitly model a vascular signal component, several of the models predicted a significant number of voxels with high D values (D > 3.1 \(\upmu\)m^{2}/ms) in one water pool. The high D values are generally less concerning when assessed in conjunction with the corresponding model likelihood maps. In Fig. 2 the very high D values seen in the anterior of the prostate in pool 2 of Model 2 mostly occur when Model 2 has a low likelihood. The high D values are much less frequent in pool 2 of Model 3 where the likelihood is high. The low likelihood of the Model 2 predictions in this region result in a low weighting in the multimodel structure prediction. In the combined ESL model prediction of Fig. 5 the predicted structure in this region is a mixture of stroma and lumen—generally consistent with the structures present in the corresponding histology sections. Nevertheless, there is possibly an excess of high D voxels, and it is unlikely that most prostates would have such a high vascular volume. Similar ‘excess vasculature’ results have been reported for VERDICT, which specifically aims to quantify the vascular component. This is an issue that requires deeper investigation, although we believe it is outside the scope of this paper.
Our ESL microstructure model assumes a Gaussian probability distribution for the contributions of individual E, S, and L components to an observed or predicted combination of water pool D and T2 values. Where available, we based our choice of probability distributions on related literature data, as detailed in Supporting Material. Other probability distributions may be more accurate, and could potentially be ascertained in a specific study that employs precision alignment of MRI and histology data. Nevertheless, the distributions defined in Table 2 produced ESL microstructure maps in generally good agreement with conventional HMMRI results from the same data^{15}. Given the uncertainties in histological validation of studies based on in vivo imaging^{30}, we cannot at this stage provide a quantitative comparison of the structure prediction accuracy of the multimodel ESL method versus conventional HMMRI.
As for the five signal models, the ESL and ’cellularity’ models are presented here only as exemplars of the flexibility of structure model assignments that is made possible when signal model parameters are not directly assigned to tissue features. While we focus on the ESL microstructure model, the proposed separation of signal modelling from tissue microstructure modelling opens the door to application of alternative biological interpretations of the signal model predictions. The twocomponent ’cellularity’ model was provided as an example of an alternative to the ESL model. The ESL model does not account for any perfusion effects from vascular water (cf. VERDICT), but this could be added as a fourth, high apparent diffusivity component with mean T2 = \(\sim\)350 ms, as expected for blood^{9}.
Where the signal acquisition protocol permits, higher order microstructure models could be applied. An obvious candidate is the addition of a third ’timedependence’ or restriction radius coordinate to the D and T2 distributions to enable characterisation of diffusionrestricting structure features^{10,13}, as used in VERDICT.
We used AICc as the information criterion for calculation of model weights. This choice is somewhat arbitrary, although model ranking based on AICc has previously been shown to correlate well with an independent test of model prediction error in a dMRI study of whole prostates ex vivo^{31}. The ’correction’ to standard AIC aims to account for situations (such as our higher order models) where number of samples (TE and b values) is low relative to number of model parameters. Other corrections (penalty terms), or information criteria such as BIC (Bayesian information criterion), might be considered. BIC tends to favour simpler models than AIC. However, discussions of best information criteria are generally focused on model selection and its implications^{32}, rather than weighting for multimodel prediction.
This paper focuses on a proposed generic method, and provides exemplars using existing MRI data. Validation of the sequential multimodel method in terms of a quantitative assessment of prediction accuracy will be highly applicationspecific, and is beyond the scope of the presented work. Rigorous histological validation, while particularly challenging^{30}, is important future work.
In summary, we have proposed a method for sequential multimodel analysis of complex MRI data where no single model accounts for both meso and micro scale tissue structure heterogeneity. In addition to implementing a method for prediction from multiple models, we suggest the proposed sequential separation of phenomenological signal modelling from structure assignment or clinical interpretation has potentially broad application in biology and other fields.
Data availability
Data are available on application to the corresponding author.
References
Epstein, J. I. et al. The 2005 International Society of Urological Pathology (ISUP) consensus conference on Gleason grading of prostatic carcinoma. Am. J. Surg. Pathol. 29, 1228–1242 (2005).
Chatterjee, A. et al. Changes in epithelium, stroma, and lumen space correlate more strongly with Gleason pattern and are stronger predictors of prostate ADC changes than cellularity metrics. Radiology 277, 751–762 (2015).
Hectors, S. J. et al. Advanced diffusionweighted imaging modeling for prostate cancer characterization: Correlation with quantitative histopathologic tumor tissue composition—A hypothesisgenerating study. Radiology 286, 918–928 (2018).
Lemberskiy, G. et al. Characterization of prostate microstructure using water diffusion and NMR relaxation. Front. Phys. 6, 91 (2018).
Bennett, K. M. et al. Characterization of continuously distributed cortical water diffusion rates with a stretchedexponential model. Magnet. Resonan. Med. 50, 727–734 (2003).
Le Bihan, D. et al. Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging. Radiology 168, 497–505 (1988).
Sehy, J. V., Ackerman, J. J. & Neil, J. J. Evidence that both fast and slow water ADC components arise from intracellular space. Magnet. Resonan. Med. 48, 765–770 (2002).
Grant, S., Buckley, D., Gibbs, S., Webb, A. & Blackband, S. MR microscopy of multicomponent diffusion in single neurons. Magnet. Resonan. Med. 46, 1107–1112 (2001).
Bottomley, P. A., Foster, T. H., Argersinger, R. E. & Pfeifer, L. M. A review of normal tissue hydrogen NMR relaxation times and relaxation mechanisms from 1–100 mHz: Dependence on tissue type, NMR frequency, temperature, species, excision, and age. Med. Phys. 11, 425–448 (1984).
Yablonskiy, D. A. & Sukstanskii, A. L. Theoretical models of the diffusion weighted MR signal. NMR Biomed. 23, 661–681 (2010).
Hall, M. G., Bongers, A., Sved, P., Watson, G. & Bourne, R. M. Assessment of nonGaussian diffusion with singly and doubly stretched biexponential models of diffusionweighted MRI (DWI) signal attenuation in prostate tissue. NMR Biomed. 28, 486–495 (2015).
Sabouri, S. et al. MR measurement of luminal water in prostate gland: Quantitative correlation between MRI and histology. J. Magnet. Resonan. Imaging 46, 861–869 (2017).
Panagiotaki, E. et al. Microstructural characterization of normal and malignant human prostate tissue with vascular, extracellular, and restricted diffusion for cytometry in tumours magnetic resonance imaging. Invest. Radiol. 50, 218–227 (2015).
Chatterjee, A. et al. Diagnosis of prostate cancer with noninvasive estimation of prostate tissue composition by using hybrid multidimensional MR imaging: A feasibility study. Radiology 287, 864 (2018).
Chatterjee, A. et al. Histological validation of prostate tissue composition measurement using hybrid multidimensional MRI: Agreement with pathologists’ measures. Abdom. Radiol. 47, 801–813 (2022).
Palombo, M. et al. Joint estimation of relaxation and diffusion tissue parameters for prostate cancer grading with relaxationverdict MRI. medRxiv 202106 (2022).
Bourne, R. & Panagiotaki, E. Limitations and prospects for diffusionweighted MRI of the prostate. Diagnostics 6, 21 (2016).
Figini, M. et al. Estimation of the vascular fraction in brain tumors by verdict correlated with perfusion MRI. In ISMRM & SMRT Virtual Conference & Exhibition 08–14 Aug 2020 (2020).
Liang, S. et al. Informationbased ranking of 10 compartment models of diffusionweighted signal attenuation in fixed prostate tissue. NMR Biomed. 29, 660–671 (2016).
Bailey, C. et al. Microstructural models for diffusion MRI in breast cancer and surrounding stroma: An ex vivo study. NMR Biomed. 30, e3679 (2017).
Ianusş, A. et al. Higherorder diffusion MRI characterization of mesorectal lymph nodes in rectal cancer. Magnet. Resonan. Med. 84, 348–364 (2020).
Fragoso, T. M., Bertoli, W. & Louzada, F. Bayesian model averaging: A systematic review and conceptual classification. Int. Stat. Rev. 86, 1–28 (2018).
Hinne, M., Gronau, Q. F., van den Bergh, D. & Wagenmakers, E.J. A conceptual introduction to Bayesian model averaging. Adv. Methods Pract. Psychol. Sci. 3, 200–215 (2020).
Brix, G., Zwick, S., Kiessling, F. & Griebel, J. Pharmacokinetic analysis of tissue microcirculation using nested models: Multimodel inference and parameter identifiability. Med. Phys. 36, 2923–2933 (2009).
Chatterjee, A. et al. Validation of prostate tissue composition by using hybrid multidimensional MRI: Correlation with histologic findings. Radiology 302, 368–377 (2022).
Gudbjartsson, H. & Patz, S. The Rician distribution of noisy MRI data. Magnet. Reson. Med. 34, 910–914 (1995).
Banks, H. & Joyner, M. L. AIC under the framework of least squares estimation. Appl. Math. Lett. 74, 33–45 (2017).
Burnham, K. & Anderson, D. Model Selection and Multimodel Inference 2nd edn. (Springer, New York, 2010).
Gordetsky, J. & Epstein, J. Grading of prostatic adenocarcinoma: Current state and prognostic implications. Diagn. Pathol. 11, 1–8 (2016).
Alyami, W., Kyme, A. & Bourne, R. Histological validation of MRI: A review of challenges in registration of imaging and wholemount histopathology. J. Magnet. Reson. Imaging 55, 11–22 (2022).
Bourne, R. M. et al. Information theoretic ranking of four models of diffusion attenuation in fresh and fixed prostate tissue ex vivo. Magnet. Reson. Med. 72, 1418–1426 (2014).
Dziak, J. J., Coffman, D. L., Lanza, S. T., Li, R. & Jermiin, L. S. Sensitivity and specificity of information criteria. Brief. Bioinform. 21, 553–565 (2019).
Acknowledgements
This work was funded under NIH grant R01 CA228036 to G.K. and A.O, with subcontract to R.B.
Author information
Authors and Affiliations
Contributions
R.B. and F.E. conceived the approach and implementation. F.E. performed the analysis and coding. A.C., G.K., and A.O. acquired the original data. T.A. performed histopathology analysis. All authors reviewed the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
EnríquezMieryTerán, F.E., Chatterjee, A., Antic, T. et al. Multimodel sequential analysis of MRI data for microstructure prediction in heterogeneous tissue. Sci Rep 13, 16486 (2023). https://doi.org/10.1038/s4159802343329x
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s4159802343329x
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.