Pattern Analysis of Dynamic Susceptibility Contrast MRI Reveals Peritumoral Tissue Heterogeneity

Abstract

Purpose

The aim of this study is to augment the analysis of dynamic susceptibility contrast MRI (DSC-MRI) to uncover unique tissue characteristics that could potentially facilitate treatment planning through a better understanding of the peritumoral region of patients with glioblastoma.

Materials and Methods

IRB approval was obtained for this study with waiver of informed consent for retrospective review of medical records. DSC-MRI data was obtained for 79 patients and principal component analysis (PCA) was applied to the perfusion signal. The first six principal components (PCs) were sufficient to characterize more than 99% of variance in the temporal dynamics of blood perfusion in all regions of interest. The PCs were subsequently used in conjunction with a support vector machine (SVM) classifier to create a map of heterogeneity within the peritumoral region and the variance of this map served as the heterogeneity score.

Results

The calculated PCs allowed near perfect separability of tissue that was likely highly infiltrated from tissue that was unlikely infiltrated with tumor. The heterogeneity map created using the PCs showed a clear relationship between voxels judged by the SVM to be highly infiltrated and subsequent recurrence. The results demonstrated a significant correlation (r=0.46, p<0.0001) between the heterogeneity score and patient survival. The hazard ratio was 2.23 (95% CI, 1.4-3.6, p <0.01) between high and low heterogeneity score patients based on the median heterogeneity score.

Conclusion

Analysis of DSC-MRI data using PCA can identify imaging variables that can be subsequently utilized to evaluate the peritumoral region in glioblastoma. These variables are potentially indicative of tumor infiltration and may become useful tools in guiding therapy as well as individualized prognostication.

Introduction

Dynamic susceptibility contrast MRI (DSC-MRI) is an important functional imaging method that enables quantitative assessment of tissue hemodynamic patterns. Aberration of blood flow, volume, and permeability is frequently observed during tumor growth and characterization of these perfusion attributes has become clinically important for both diagnosis and therapy planning. In the context of glial neoplasms, perfusion characteristics have been shown to correlate with tumor type and grade (1), and hence influence treatment decisions. Glioblastoma is the most common and aggressive glial tumor in adults (2) with abnormal (tumor) tissue extending far beyond the visible, enhancing tissue on MRI (3). However, prevailing methods for perfusion data analysis yield little, if any, additional information in the surrounding peritumoral region. Furthermore, although measurements such as relative cerebral blood volume (rCBV), the most common method for analyzing and interpreting data obtained from DSC-MRI, have shown promising correlations with both histologic grade and degree of neovascularization in gliomas (4, 5), certain tissue characteristics and perfusion dynamics may not be fully captured by such global, one-dimensional variables (scalars).

DSC-MRI is based on the principle that flow of a paramagnetic contrast agent through a capillary bed will transiently change the magnetic susceptibility of the given tissue (6). Decreased signal on spin or gradient echo sequences after first pass of the contrast agent, frequently described as susceptibility induced T2* shortening, is the result of this temporal change in magnetic susceptibility. This signal time-curve is then converted into a concentration time-curve and using tracer kinetic analysis various hemodynamic variables such as cerebral blood volume (CBV), blood flow (CBF), and mean transit time (MTT), as well as metrics addressing vessel leakage such as percent signal recovery (PSR), may be estimated (7). Combined, these metrics enable microvascular imaging, providing a visual correlate of blood flow, volume, and vessel permeability (8, 9).

However, calculation of these metrics is not only complicated but also involves various assumptions and model-dependent approximations (7). For example, bolus tracking relies on the assumption that the endothelial membrane is intact and there is no leakage of contrast agent, while fitting of a gamma-variate function to the concentration time-cure is required to correct for tracer recirculation (10). Additionally, in order to calculate quantitative perfusion variables, the arterial input function, the concentration of contrast agent as a function of time, is frequently estimated from voxels near major arteries (7, 10). Partly due to the post-processing complexities involved with DSC-MRI, commonly used variables in clinical use often describe the breadth of the perfusion signal with only a couple of variables such as rCBV.

We hypothesize that additional, complementary features may be extracted from DSC-MRI when the entire perfusion signal is analyzed as a time-series, using methods that capture all characteristics of the shape of this signal. The importance of the complete perfusion time-series for tissue characterization has been previously highlighted. Chou et al. demonstrated that noiseless independent factor analysis may be applied to DSC-MRI data in order to extract spatiotemporal blood supply patterns of various tissue compartments (11). In turn, these patterns may be used to systematically classify tissue and improve the differentiation between normal and abnormal hemodynamics in a given tissue type.

In this work we use dimensionality reduction methods to analyze the perfusion time-series of patients with glioblastoma in order to identify tissue features that are not captured by currently calculated variables (e.g. rCBV). To investigate the clinical utility of these multidimensional features, we employ machine-learning tools to analyze the heterogeneity of the peritumoral region. This region is an important tissue area in glioblastoma that leads to tumor recurrence in over eighty percent of patients (12). The aim of this study is to augment the analysis of DSC-MRI to uncover unique tissue characteristics that could potentially facilitate treatment planning through a better understanding of the peritumoral region of patients with glioblastoma.

Materials and Methods

Data collection and preprocessing

Institutional review board approval was obtained for this study with waiver of informed consent for retrospective review of medical records. Patients treated at our institution for glioblastoma (WHO Grade IV) between July 2006 and December 2011 were retrospectively selected for our study. The inclusion criteria was a patients with sole tumor with pre-operative advanced MRI (anatomic and perfusion imaging modalities). Subjects who had a prior tumor treated by radiation, surgery, or chemotherapy were excluded from the study. 9 patients were excluded due to multiple tumors, 8 patients were excluded due to missing perfusion MRI, and 6 patients were excluded due to prior resection. This resulted in 79 de novo patient samples, all of which were treated under the same protocol. The clinical diagnosis of tumor recurrence was established via histopathological report after re-operation to exclude pseudoprogression (13). All patients were imaged using a 3T MR imaging scanner (Magnetom TrioTim; Siemens) and the image acquisition protocol was constant. With regards to the DSC sequence, the parameters were as follows: FOV: 22cm, Matrix: 128 by 128, Pixel Spacing: 1.72 by 1.72mm, Slice thickness: 3mm, TE/TR: 45/2000ms. An initial loading dose of 3mL of MultiHance (gadobenate dimeglumine) was administered, which after five minutes was followed by another bolus injection with the remaining dose (for a total of 0.3mL/kg or 1.5 times single dose) during image acquisition.

All MRI studies underwent the following pre-processing procedures (by XD, medical image analyst, 6 years experience). SUSAN (14), a low level image processing method, was used for noise reduction and the nonparametric nonuniform intensity normalization (N3) (15) algorithm was applied to correct for MRI intensity nonuniformity. Subsequently, image alignment was performed using affine registration (FLIRT) (16) and skull stripping was carried out using the brain extraction tool (17) of FSL (18). Finally, GLISTR (19), a technique that combines biophysical models of tumor growth with image-based tissue modeling, was used to create a segmentation mask of the enhancing tumor as well as the peritumoral, edematous tissue.

Calculation of Principal Components

Principal component analysis (PCA) is a standard dimensionality reduction method (20) that was used to distill the 45 second DSC-MRI time series down to a few variables (principal components) that capture the temporal dynamics of blood perfusion. In particular, the first principal component is a projection of the signal onto a direction that captures the highest amount of variance that can be captured by a single variable. Each succeeding component in turn has the highest variance possible under the constraint that it be uncorrelated with the preceding components. These components concurrently present various aspects of the shape of the perfusion time-curve. Each voxel’s perfusion was sampled at 45 time points which were then used as voxel’s feature vector in the PC analysis, yielding 45 eigenvalues and correspondingly 45 components. We retained the number of components that accounted for more than 99% of the overall variance in the perfusion signal across individuals, indicating that the few PCs retained for further analysis captured almost all the perfusion dynamics present in the DSC-MRI signal.

In order to characterize the perfusion characteristics of different brain tissues, several regions of interest (ROIs) were defined (white matter, gray matter, CSF, edema, enhancing tumor, and non-enhancing tumor). Additionally, two ROIs were placed within the edema/peritumoral region, one immediately adjacent to enhancing tumor (near) and the other at the distal edema boundary (far) (by HA and LM by agreement, LM, Neurosurgery resident, PGY-6 and HA, international medical graduate and medical imaging expert, PhD, 9 years experience). The maximum width of these ROIs was two voxels to maintain region homogeneity and specifically did not include any area that was contrast-enhancing. These two ROIs served as reference examples for near-tumor and far-from-tumor tissue, based on the expectation that they are likely to have relatively higher and lower infiltration respectively (21, 22), albeit they are merely modeling parameters and by no means imply anything about true underlying (unknown) infiltration.

The mean perfusion curve was computed for each ROI based on the perfusion signal time-curve of all voxels in that ROI (Figure 1, left panel). PCA was subsequently used to capture the information of the perfusion time series in all ROIs and all subjects. The projections of data onto principal components contain the new coordinates of the data points. Because of the relative consistency in the perfusion pattern of the various ROIs, a feature vector consisting of first six principal components was sufficient to capture more than 99% of the variance in the perfusion signal for all tissue types and all patients (Figure 1, right panel, depicts the first three components).

Figure 1. Perfusion Time-Series and Calculated Principle Components.

The left panel shows the average perfusion signal (Dynamic susceptibility contrast MRI time-series) of all voxel for a given region of interest (ET: enhancing tumor; NCR: non enhancing core; ED: edema; GM: gray matter; WM: white matter; CSF: cerebrospinal fluid), while the right panel demonstrates the calculated principal components for each tissue type, based on the perfusion signal. PC1, PC2, and PC3 denote the first, second, and third principal components, respectively.

The conventional way to represent the perfusion signal is via Delta-R2*, which is calculated by dividing the signal to the value of the baseline and obtaining its logarithm to convert the signal into concentration-time curves for single-echo data. Herein, however, our premise is to use machine learning methods to characterize the heterogeneity of the perfusion signal in edema, with the goal of predicting infiltration, recurrence and survival. The transformations involved in calculating Delta-R2* might therefore not be optimal from this perspective and might remove (by removing baseline) or distort (by taking the logarithm) useful information from the perfusion signal. We therefore took the machine learning angle, and used the raw MR signal instead. However, the PCA was applied to Delta-R2* as well to numerically prove the superiority of using raw data (Table 1).

Table 1.

Correlation Analysis

Variable	Correlation (r)	Significance (p value)
Survival Analysis
Heterogeneity score	0.46	0.00002
rCBV	0.04	0.72
Delta-R2*	0.35	0.002

The correlation between perfusion imaging variables and survival is shown. (rCBV = relative cerebral blood volume)

Evaluation of Principal Components

The principal components (PCs) incorporate different aspects of the dynamics of blood perfusion, including baseline signal, depth of signal drop, slope of signal drop and recovery, relative recovery, etc. (refer to Figure 1, left panel). Hence, they capture and convey complex, high dimensional information about a given tissue type as compared to the currently calculated scalars, such as rCBV. To investigate whether these PCs carried additional information, we focused our analysis on the peritumoral region. As previously noted, in this region current perfusion measures fail to convey any useful information.

To accomplish this, we used support vector machines (SVM) (23) to interpret the principal components of all voxels in the peritumoral region of a given patient. SVM is a learning model that analyzes data and recognizes patterns, used for classification. It takes a set of input data and predicts which of two possible classes forms the output. SVM was trained to create a classifier that, using the six PC input features, aimed to distinguish between edematous peritumoral tissue immediately adjacent to the enhancing tumor and edematous tissue further away. This schema was designed based on the assumption that the former is likely to be relatively more highly infiltrated than the latter. In other words, based on the PCs of a given voxel, the SVM produced a score that was positive if the perfusion dynamics were similar to those in the low-infiltration ROI; in turn, a negative score was generated when the perfusion dynamics were similar to that of the high-infiltration ROI. This was performed on a voxel-by-voxel basis for the entire peritumoral region (by HA).

The SVM was trained using libSVM (24) with a Gaussian kernel function, and parameters were optimized based on a cross-validated grid search. Subsequently, the SVM model was trained using the projections of the near-far ROIs and applied to the rest of the peritumoral region. This process was repeated for all subjects and the generated scores were used to produce a map of heterogeneity within the peritumoral region. The variance of these scores in the peritumoral region defined the heterogeneity score for a given test subject (by HA). The heterogeneity score and the variance of rCBV within the peritumoral region were used to perform a correlation analysis between imaging variables and patient survival.

We evaluated the cluster heterogeneity to determine if this provided additional, complementary information about a given tissue type beyond currently used scalar measures. For this analysis we focused exclusively on the edematous, peritumoral region. The SVM that was trained using the high and low infiltration ROIs described earlier was applied to the entire peritumoral region. Resultant SVM scores were used to generate a “color scale” of heterogeneity within the peritumoral region.

In the absence of a correlate tissue sample from the disparate regions, we performed a qualitative comparison with the post-recurrence MRI to assess what additional information these maps of heterogeneity carried. For this study, only patients that had no obvious residual after initial resection based on all modalities including rCBV and had pathology proven tumor recurrence were evaluated (by MB, Neuroradiologist, 12 years experience, LM, and HA by agreement).

All statistical analyses (correlation, Hazard ratio, Kaplan–Meier curves) were performed using a statistical package (SPSS Version 21; Armonk, NY: IBM Corp) (by XD). Correlation between imaging variables and survival was obtained by Pearson’s correlation. Subsequently, the subject group was dichotomized into high and low heterogeneity on the basis of the median heterogeneity score and hazard ratios and Kaplan–Meier curves were computed for survival analysis, with the level of statistical significance for a two-sided comparison set at 5% (p value < 0.05).

Results

DSC-MRI Time Series

Five distinct periods were observed in the time series of DSC-MRI image data, consistent with theoretical models and previous work (Figure 1) (10). Initial inspection of the perfusion time series in Figure 1 reveals that different brain tissues have unique perfusion dynamics. For example, the enhancing tumor region is, on average, associated with the largest signal drop, while the peritumoral region has the smallest drop in signal. The various ROIs also differ in their baseline signal characteristics. Regions of edema and CSF have the highest baseline signal, while white and grey matter have the lowest. This is not unexpected given that the DSC protocol is a T2* weighted sequence.

Evaluation of Perfusion Time Series through Principal Components

When the perfusion time-series are evaluated using principal components, similar brain tissue forms characteristic clusters. These clusters, although heterogeneous, define a specific ROI (e.g. non-enhancing tumor, Figure 1, right panel) and are separable from other tissues (clusters). More importantly, the calculated principal components of the high and low infiltration ROIs allow clear separability of these two important regions. Figure 2 illustrates the probability density function of the high and low infiltration ROIs, as calculated using the described method (PC-SVM score) as well as commonly computed perfusion measure (rCBV). While our method leads to two distinct, non-overlapping histograms, the remaining perfusion measures do not afford the same level of tissue separability.

Figure 2. Separability of Voxels within Peritumoral Region.

The figures demonstrate the separability of the near (likely infiltrated) and far (not infiltrated) region of interests (ROIs) within the peritumoral region. Red represents the probability density function of the near voxels, while blue represents the far voxels. The first panel, based on our method, shows two completely separable histograms (groups). The right panel based on relative cerebral blood volume (rCBV) intensity values shows overlapping histograms (X-axis shows the intensity in arbitrary unit scaled between 0-255 and Y-axis is the number of voxels).

The left panel in Figure 3 illustrates this color scale map for two representative subjects. Red areas depict regions that are most similar to the high infiltration ROI and blue depicts regions that are similar to the low infiltration ROI. Images were registered using affine registration (FLIRT), as outlined in the methods. In a large majority of patients, the region of recurrence was in close proximity to the area judged by the SVM to be similar in perfusion characteristics to the high infiltration ROI (See Figure 3).

Figure 3. Map of Peritumoral Heterogeneity.

Images from two representative patients are illustrated after PC analysis. Left to right, the first column shows the preoperative T1 post gadolinium the second column shows the rCBV map for each subject, which the peritumoral region demonstrating relatively low, homogeneous perfusion. The third column shows the map of heterogeneity generated from the SVM scores within the peritumoral region. Red regions are areas most similar to highly infiltrated tissue, while the blue regions are more similar to low infiltrated tissue. The last column depicts the T1 with gadolinium sequence, at the same slice, for each subject after tumor recurrence.

The results demonstrated a high and significant correlation (r=0.46, p<0.0001) between the heterogeneity of the peritumoral region and patient survival, while the score using Delta-R2* had lower correlation (r=0.35). On the contrary, the rCBV did not correlate with survival (see Table 1). Figure 4 illustrates the Kaplan–Meier survival curves for high and low heterogeneity groups respectively (Hazard ratio 2.23, 95% CI 1.4-3.6, p <0.01). Finally, support vector regression analysis demonstrated that the first six PCs capture the information conveyed by common perfusion measure, rCBV (r=0.94).

Figure 4.

The figure shows the Kaplan-Meier survival curve for the three subject groups, low (red color), medium (orange color), and high (green color) heterogeneity respectively. The analysis was based on 79 subjects that were dichotomized according to the heterogeneity score. The calculated Hazard ratio for low heterogeneity group is 2.23 (95% CI 1.4 – 3.6, p < 0.01).

Figure 5 provides a visual representation of the first four principal components extracted from the DSC-MRI perfusion sequence. In Figure 5 (right panels), the first four principal eigenvectors obtained from all voxels are plotted after being multiplied by +/- two standard deviations of the respective principal components. These plots indicate that the different principal components were found to relate to different aspects of the perfusion curves: 1) PC1 was found to primarily relate to the global baseline signal level at each voxel; 2) PC2 has the greatest variability around the baseline and the depth of the curve, therefore it conveys the depth of the signal drop in relation to the baseline level; 3) PC3 reflects more complex information, namely the shape of the perfusion signal, in other words how steep the signal drop and recovery are; 4) PC4 appears to be driven by the baseline signal and its recovery.

Figure 5. Principal Component Images and Plots.

The first four principal components of an MR image (left), along with the plot of the corresponding principal eigenvector (right), are shown to illustrate the breadth of information contained within the perfusion time-series. The plots have been constructed from the perfusion signal of all voxels, with the error bar representing +/− 2 standard deviations of the respective principal component. Red and blue lines represent the negative and positive parts of PCs respectively.

Discussion

The proposed method highlights the heterogeneity of the peritumoral region using the temporal dynamics of DSC-MRI. This heterogeneity map, which is constructed using advanced but easy to implement analytical techniques, demonstrates the clinical relevance of this region. Notably, we found that the heterogeneity within this region is significantly, and robustly, correlated with patient survival and offers insight into potential future tumor recurrence. This information is completely obscured when the perfusion signal is analyzed in the traditional manner. Thus, this methodology not only highlights the importance of the peritumoral region in glioblastoma, but also how information from this region, when analyzed in this manner, can bring to light important clinical information that may otherwise have been missed.

The peritumoral region remains a critical problem both in the understanding and treatment of glioblastoma. Although it has been previously shown (3, 26) that edema results from infiltrating tumor cells, but edema also occurs as a biological response to the angiogenic and vascular permeability factors released by the spatially adjacent tumor cells (27). The proposed method extracts information that is currently obscured when analyzing perfusion images using prevailing techniques and our preliminary results demonstrate that this information (tissue heterogeneity) may reflect tumor invasion, as evidenced by the close spatial relationship with tumor recurrence. These results are strengthened by the fact that our approach does not require any assumptions or approximations, and, more importantly, are clinically relevant as substantiated by the strong correlation between our calculated heterogeneity score and patient survival. Thus, although DSC-MRI without a correlate pathology specimen cannot discern the heterogeneity of the peritumoral region, our calculated heterogeneity score can be used to differentially classify this important tissue region.

A limitation of this study is lack of a biological reference index such as histopathology correlation with surgical specimens from the peritumoral infiltrated areas that would be addressed in our future study. Another limitation of our approach relates to the specialized nature of PCA and SVM, and the lack of availability of related software on regular clinical workstations. However both PCA and SVM are widely understood and described procedures, with lots of free software available for their application. Moreover, our own software pipeline is freely available for use.

In summary, advanced imaging techniques are increasingly used in the clinical evaluation of human gliomas. DSC-MRI has already been applied to differentiate glioblastoma from brain metastasis, predict glioma grade, and distinguish recurrence from radiation necrosis (26). The present study extracts informative features from the temporal dynamics of DSC-MRI using principal component analysis, and employs these variables via SVM classification to highlight the heterogeneity of the peritumoral region. Although preliminary, this method may potentially identify highly malignant regions that would have otherwise not been recognized using current techniques. The results of this study represent the methods for analyzing the MR perfusion signal that enable improved characterization of the peritumoral region as well as localization of highly infiltrated areas. In turn, this information may be used to augment targeted therapy and provide patient specific prognostication.

Advances in Knowledge.

• Principal component analysis (PCA) of perfusion time-series, coupled with machine learning methods, offer a mathematically rigorous way of quantifying the full dynamics of blood perfusion in dynamic susceptibility contrast MRI (>99%).

• These comprehensive measurements go beyond standard relative cerebral blood volume (rCBV) and may be utilized to extract peritumoral imaging features highlighting underlying subtle tissue heterogeneity (p<0.0001).

Implications for Patient Care.

• The method proposed in this paper enables clinicians to visualize the heterogeneity of the peritumoral region in glioblastoma patients and appreciate regions of subtly different imaging characteristics that would otherwise be visually missed.

• Imaging indices derived via this approach seem to correlate with survival and recurrence, and therefore can be used as predictive tools.

Summary statement.

The present study extracts informative features from the temporal dynamics of DSC-MRI using principal component analysis, and employs these variables via SVM classification to highlight the heterogeneity of the peritumoral region.