Research ArticleCancer

Metabolomic Imaging for Human Prostate Cancer Detection

See allHide authors and affiliations

Science Translational Medicine  27 Jan 2010:
Vol. 2, Issue 16, pp. 16ra8
DOI: 10.1126/scitranslmed.3000513


As current radiological approaches cannot accurately localize prostate cancer in vivo, biopsies are conducted at random within prostates for patients at risk for prostate cancer, leading to high false-negative rates. Metabolomic imaging can map cancer-specific biomolecular profile values onto anatomical structures to direct biopsy. In this preliminary study, we evaluated five whole prostates removed during prostatectomy from biopsy-proven cancer patients on a 7-tesla human whole-body magnetic resonance scanner. Localized, multi–cross-sectional, multivoxel magnetic resonance spectra were used to construct a malignancy index based on prostate cancer metabolomic profiles obtained from previous intact tissue analyses with a 14-tesla spectrometer. This calculated malignancy index is linearly correlated with lesion size and demonstrates a 93 to 97% overall accuracy for detecting the presence of prostate cancer lesions, suggesting the potential clinical utility of this approach.


The prostate-specific antigen (PSA) blood test has greatly improved the detection of human prostate cancer at early and asymptomatic stages of cancer development (13). However, PSA values, although prostate-specific, are not cancer-specific, and benign prostate conditions may elevate PSA readings. Thus, at present, prostate cancer can only be definitively diagnosed with a positive biopsy. Moreover, ultrasound-guided prostate biopsies now ensure only that the biopsies are from somewhere in the prostate; they unfortunately cannot be selectively obtained from cancer-suspicious locations. Because of the extremely heterogeneous distribution of prostate cancer, such biopsies have high false-negative rates (4). New radiological paradigms that could indicate suspicious regions for selective biopsy before prostatectomy would alleviate many difficulties and controversies now associated with the clinical management of this disease (5).

Until recently, magnetic resonance (MR) images for detection, localization, and staging of prostate cancer were largely based on T2-weighted studies. However, advanced imaging technologies, such as functional MR imaging (MRI) and MR spectroscopy (MRS), have been increasingly used in diagnostic radiology (6) to address limits in the clinical utility of T2-weighted images. Such functional MRI technologies as apparent diffusion coefficient (ADC) maps from diffusion-weighted images, Ktrans maps from modeling of dynamic contrast enhancement (DCE) data, and metabolic ratios from MRS are now being investigated to improve diagnostic capability (7). Still, these individual MR approaches reveal only selected parameters associated with cancer; none yet can reveal the full extent of prostate cancer. A clear need thus exists for multiparametric imaging that can assess complicated disease processes.

At present, the production and interpretation of quantitative multiparametric MRI maps aimed at improving differential diagnosis in the context of malignancy must involve the collaborative contributions of expert radiologists and MR scientists. Efforts are under way, however, to render findings from multiparametric data sets obtained with MRI more relevant to the clinical context and more specific for parameters of interest. For example, combining T2-weighted images with ADC maps can help predict prostate cancer volume (8), whereas combining them with DCE-MRI can improve the sensitivity and specificity of cancer detection for patients who have previously had negative biopsies, but who are at high suspicion for cancer (9). The relation of MRS metabolic ratios to tumor stages has also been explored (10, 11). However, the particular contribution of MRS to cancer diagnosis, as distinct from other MR approaches, lies in its ability to evaluate cellular biology and chemistry, rather than the physical properties of water in tissue; this ultimately could permit a routine objective test that could facilitate personalized prostate cancer treatment (1215).

Our recently developed capacity to assess molecular and physiological parameters at improved resolutions, as well as to exploit statistical approaches derived from medical genomics and proteomics, has made it possible to evaluate, for the purpose of clinical diagnosis, treatment strategy, and prognosis, an entire ensemble of measured parameters rather than analyzing these factors individually. Here, instead of investigating individual metabolites or their ratios, as would be done for medical MRS or MRS imaging (MRSI), we have exploited the existence of relations between cellular metabolism and specific pathological conditions to investigate the entire ensemble of measured prostate metabolites—metabolomics—so as to determine whether tumors can be localized by imaging prostate cancer–specific metabolomic profiles.

Although particular metabolic pathways may be more directly linked with certain diseases, the collective evaluation of all measurable metabolic changes could more fully characterize individual pathological conditions (1618). On the basis of this rationale, we previously hypothesized that simultaneous global evaluation of all measurable metabolites may reflect underlying pathological conditions more accurately than can the analysis of any single metabolite, a possibility assessed in a “bench-top” pilot study of intact human prostate cancer tissues, conducted on a 14-T MR spectrometer with a high-resolution magic angle spinning (HRMAS) technique (19). HRMAS exploits solid-state NMR’s magic angle spinning technology (20, 21), generates detailed data of tissue metabolic compositions, and preserves the tissue pathological architectures needed to correlate metabolic changes with pathological alterations accurately (22). In our previous pilot study, we established parameters of metabolomic profiles from intact tissue analyses that identified the presence of prostate cancer and indicated tumor pathological stages. We also demonstrated that these metabolomic profiles showed a greater sensitivity and specificity in identifying pathology than did individual metabolites (19).

To assess and extend this earlier observation, we hypothesized that diagnostic radiology’s capacity to locate malignancy in prostates could be improved through the imaging of prostate cancer metabolomic profiles, instead of individual metabolites as now practiced in MRSI (23). We tested both our previous and extended hypotheses by examining excised human prostates with a 7-T human whole-body MR scanner that could be integrated into the clinical armamentarium, with prostate cancer metabolomic profiles obtained from previous analyses of intact tissue on the 14-T MR spectrometer.


Construction of metabolomic profiles for phantom imaging with principal components analysis

Despite its conceptual simplicity, the translation of measured metabolite intensities from intact tissue MRS into profiles capable of characterizing specific pathologies presents a practical challenge. Even before considering pathologies, the complex metabolomic data matrix must be simplified through dimension reduction mechanisms; this was achieved statistically by principal components analysis (PCA). PCA defines orthogonally independent PCs that present as linear combinations of measured metabolites so that the first of a few PCs cover most data variations of the entire matrix. Thus, the evaluation of relations between all metabolites and the pathology of interest reduces to an analysis of potential correlations between a much smaller set of PCs and this pathology.

To illustrate the PCA procedure and test the validity of the prostate cancer metabolomic profiles (obtained from intact tissue MRS at 14 T with HRMAS) for prostate cancer imaging of the whole prostate at 7 T with multivoxel MRS, we conducted a phantom study that evaluated this transition between differing experimental conditions. Three solutions of common biological metabolites were prepared, including creatine, phosphocreatine, l-glutamic acid, sodium citrate, l-glutamine, taurine, myo-inositol, N-acetylasparate, choline, phosphorylcholine, glycerophosphorylcholine, lactic acid, acetate, spermine, spermidine, and putrescine. The concentrations of these metabolites were prepared by random variations of each chemical between 5 and 20 mM of physiological levels. To imitate intact tissue analysis, we measured four independent spectra of 10 μl with HRMAS at 14 T for each of the three solutions (Fig. 1A). From these spectra, the 10 most intense resonance peaks were selected and analyzed with PCA. Both PC1 and PC2 showed statistical significance in discriminating among the three solutions, with coefficients representing each measured metabolite in the linear combination for PC2 (the so-called loading factors of PCA; Eq. 2 in Materials and Methods) presented as an illustration (Fig. 1B).

Fig. 1.

Metabolomic imaging of solution phantoms. (A) Spectra of three solutions made from common biological metabolites at varying concentrations (5 to 20 mM; see text for details). The bottom traces are spectra acquired at 14 T with the same HRMAS procedure as that used for intact tissue analysis, whereas the top traces are spectra measured from voxels indicated in (C). (B) Coefficients for PC2, as a metabolomic profile calculated from the 10 most intense metabolic resonances measured at 14 T. (C) Image for the placement of multivoxel MRS at 7 T. (D) Values of metabolomic profiles measured from voxels can differentiate between three solution phantoms with statistical significance (analysis of variance) (P < 0.0001). (E) The intensities of voxel profile values are color-plotted according to the color-coding bar at the right.

We then situated three solution-filled spherical phantoms (ϕ = 19 mm) in one inclusive gel phantom (ϕ = 48 mm) of a size comparable to that of a human prostate, positioned it within a 7-T human whole-body MR scanner, and measured localized multivoxel MRS (Fig. 1C). The resulting voxel spectra of corresponding solutions were compared with the measurements at 14 T (Fig. 1A). PC2 values, calculated with metabolite coefficients (Fig. 1B), could differentiate among the three spherical phantom solutions with statistical significance (P < 0.0001) (Fig. 1D). The PC2 color map provides a visual presentation of this differentiation (Fig. 1E). These phantom PCA results supported the applicability, as a first step of approximation, of prostate cancer metabolomic profiles obtained with intact tissue at 14 T to the imaging of whole prostates at 7 T.

Identification of prostate cancer metabolomic profiles through canonical analysis

Because no external concentration references can be effectively established for the entire measured whole prostate, we had to reanalyze our previously established, concentration-based, prostate cancer metabolomic profiles (19) according to their relative metabolite intensities, with each individual metabolite intensity normalized by intensity of the total metabolite region of 0.5 to 4.5 parts per million (ppm) as measured from the same spectrum. Similar to published protocols, these reanalyses used both PCA to reduce data dimensions and canonical analysis to reorient the spatial distributions of PCs, chosen for their significant correlation with tissue pathology. This reorientation extends beyond the PCA requirement for orthogonality among PCs and recognizes simultaneous and independent relations of multiple PCs with a particular pathological event. The detailed equations representing scores from this canonical analysis are presented in Materials and Methods.

After the reanalysis of PCA with the relative intensities of all 36 metabolite regions identified in our previous study (19), a new metabolomic profile, canonical score 1, was calculated by selecting PC1, PC3, PC6, and PC10, each of which, with statistical significance, either was linearly correlated with volume percentage of prostate pathology features or could differentiate cancer from histologically benign samples (Table 1). Overall loading factors (f in Eq. 3, Materials and Methods; this variable combines coefficients from both PCA and canonical analysis) for the evaluated 36 metabolite regions were obtained for standardized relative intensities (Fig. 2A).

Table 1.

Characteristics of the four PCs involved in canonical analysis. The first 10 PCs were evaluated against tissue pathologies on the basis of the criterion of eigenvalue ≥ 1.0. VEigen, eigenvalue; Cum%, cumulated percentage; CaP, prostate cancer; HB, histologically benign; Epi, epithelia.

View this table:
Fig. 2.

Prostate cancer metabolomic profiles of relative metabolic intensities at 14 T. To compensate for the lack of an established in vivo concentration reference standard, we reanalyzed tissue metabolomic profiles according to relative metabolite intensities (normalized by the metabolite spectral region of 0.5 to 4.5 ppm) for 42 samples from 13 patients (19). (A) The overall loading factors (combined coefficients from PCA and canonical analysis) for the 36 metabolites and regions included, which provide examples of phosphorylcholine (PCh; 3.22 ppm), spermine (Spm; 3.05 to 3.15 ppm), and creatine (Cr; 3.03 ppm), are labeled. (B) Metabolomic profile expressed as canonical score 1 distinguished cancer (solid dot) from histo-benign (open dot) samples (overall accuracy of 93%, indicated by the ROC curve, not shown here) with statistical significance (P < 0.0001). Median (M) and SD values were calculated for all samples.

The metabolomic profile calculated from relative intensities was slightly less significant [overall accuracy of 93%, as presented by the area under the curve calculated from the receiver operating characteristic (ROC) curve] than the one previously reported for metabolite concentrations (overall accuracy, 98%) (19). However, the new profile was able to differentiate cancerous samples (15) from histologically benign ones (27) collected from 13 patients [Gleason score (GS) of 5 (one case), 6 (seven cases), 7 (four cases), and 9 (one case)]. The median (M, black square) and SD (vertical bar) of the 42 tested samples were determined (Fig. 2B). The values of all histo-benign samples fell below the empirical and arbitrary threshold median plus one SD (M+SD). Of note, 12 of the 13 patients whose data are included in this figure had pathological stage II (T2) tumors (disease confined in the organ), as defined by the American Joint Commission on Cancer tumor/lymph nodes/metastases (TNM) classification (5); one was a T3 (tumor extended beyond the prostate) patient of GS 7. Thus, this relative resonance intensity–based metabolomic profile, although able to differentiate cancer from histo-benign tissue, may be tumor stage–sensitive and heavily skewed toward T2 tumors.

Metabolomic imaging of excised human prostates

Five prostates, freshly removed by prostatectomy, were measured with the 7-T MR scanner. To avoid any interference with clinical pathology after MRS measurements, we acquired spectra with these prostates placed in air, without immersing them in fluids of magnetic susceptibility compensation. A single-loop coil was placed axially around the middle of the prostate, and multivoxel MR spectra were acquired for three prostate cross-sectional planes, with the center plane situated approximately at coil level, and two parallel off-center planes 6 mm above and below the center plane. Each plane (3 mm thick) consisted of 16 × 16 voxels, each ~3 mm × 3 mm, as prescribed on a T2-weighted axial image (Fig. 3A).

Fig. 3.

Metabolomic imaging of cancer from excised human prostate. (A) MRI of a prostate axial cross section (from a 47-year-old TNM stage pT2cNxMx patient) overlaid with a grid to indicate the locations of 16 × 16 voxels for which multivoxel MR spectra were acquired. The outer white border delineates the outline of the prostate cross section; the inner white border circles the urethra. Spectra in voxels outside the outer border and inside the inner border were not included in the analyses. (B) Because of the magnetic susceptibility interference at the tissue-air interface, spectra in voxels between solid and dashed lines were eliminated from further analysis. The values of the metabolomic profile for the remaining voxels were calculated with coefficients presented in Fig. 2A (Eq. 3) for all remaining voxels; the values were then mapped onto the MR image with a color range calibrated to −100 and 100. (C) Histologically identified cancer regions are circled in red and plotted onto a whole-mounted prostate histology image at approximately the same prostate cross-sectional level as in (B). Metabolomic profile regions having at least two connected voxels with profile values greater than M+SD are plotted in shaded red. Partial tumor sizes are estimated by excluding their overlaps with purple voxels; here, the “partial tumor size” for the right lesion on the top right of the histological slide equals the total tumor size, whereas for the cancer region on the top left of the histological slide it is less than half the tumor size. Geometrical centers are used to calculate the center of profile-elevated regions and to estimate the center of partial tumor size, excluding the discarded voxels. (D) Representative spectra from voxels with elevated [a in (B)] and nonelevated [b in (B)] profile values are plotted.

Spectral data from voxels within the white border, excluding the three voxels of the urethra (H8, I7, and I8), were processed individually. Spectra from the voxels near the prostate edges were too featureless to be included for further analysis, likely because of susceptibility effects at the tissue-air interface, and were excluded. For each remaining voxel, we applied the overall loading factors of each metabolite region, obtained from 14-T data (Fig. 2A), on metabolic intensities normalized by the spectral intensity between 0.5 and 4.5 ppm from the same voxel, to calculate a prostate cancer metabolomic profile value for the voxel. We individually examined the calculated values for voxels adjacent to the excluded voxels to prevent any “out-of-voxel” interference from affecting the spectra of voxels adjacent to the excluded, featureless edge voxels.

Our prostate cancer metabolomic profiles were structured to produce a higher value wherever voxels appeared more cancer-suspicious (Fig. 2B). For this reason, if the profile value of an adjacent voxel was higher than its immediate inner voxel, the value of the particular adjacent voxel was excluded and the whole data set was recalculated to avoid potential false positives caused by out-of-voxel susceptibility effects. After exclusion of all potentially contaminated voxels, the final calculated profile value for each voxel was weighted by the values of surrounding voxels through a deconvolution process. Finally, all resulting values of the metabolomic profile were overlaid onto the MR anatomic image (Fig. 3B), with excluded voxel areas marked between solid and dashed lines. These analyses and exclusions were conducted before reviewing the pathology.

Correlation between metabolomic images with cancer histology

The results of whole-mounted pathology obtained at multiple levels of the five prostates revealed that the center planes of two prostates contained no histologically detectable cancerous lesions. For each of these two prostates, both “above and below” off-center planes were included in the analysis to reduce any potential bias that might have been introduced by the distance between an off-center plane and the coil level. Thus, for these five prostates, seven 16 × 16 voxel cross-sectional planes were analyzed.

The histopathological image of the prostate at approximately the same level as the cross-sectional plane examined by the MRS revealed irregular cancer lesions in the top left and right corners (Fig. 3C). By calculating the value of M+SD, established as a testing threshold (Fig. 2B), we identified voxel areas that had at least two connected voxels, each having a profile value (that is, the calculated canonical score 1 obtained based on Fig. 2A) higher than M+SD, as determined for the particular cross-sectional plane. Representative spectra from elevated and nonelevated voxels are shown in Fig. 3D.

Independently from the results of histopathology, we identified, from the seven analyzed cross-sectional planes, 13 profile-elevated regions, each of which had at least two consecutive voxels. All these voxels have values of metabolomic profiles above the threshold of M+SD, as determined for their corresponding planes. However, because M+SD values are calculated within each examined cross-sectional plane, and certain voxels in any given plane will have values higher than their own M+SD, whether or not prostate cancer is present, the M+SD values are relative and difficult to cross-examine among different planes and for different cases. To circumvent this relativity, we defined a malignancy index for these cross-sectional planes (MIp) as follows:MIp=PV>M+SDi=1nPVi×VSi(1)where the sum represents the addition of all connected voxels that have prostate cancer profile values higher than the threshold, with PV representing profile value and VS representing voxel size.

Overall, this index measures the absolute elevation of the metabolomic profile and may be compared among different cross sections from different cases. The seven cross sections analyzed for the five prostates yielded seven histologically identifiable tumor regions in five cross sections. Of the seven histologically identified tumor regions, two were located entirely in the excluded voxels and thus could not be included for further analysis. For the remaining five tumors (all GS 7, four T2 and one T3), we estimated two-dimensional size (mm2) from each histological image, as well as partial size, by excluding the areas overlapped with the excluded voxels. We then identified five metabolomic regions from the above-mentioned 13 profile-elevated regions that potentially corresponded to a histological region of the five tumors, respectively. The identified region, on the same plane as the tumor, either overlapped with the histological tumor region or was located at less than the length of one voxel between the edges of histological and metabolomic regions. The remaining eight regions, which either were situated within prostate imaging planes that had no histologically identifiable tumor or were more than two voxels away from the tumor on the same plane, were considered as histo-benign, although with elevated metabolomic profile values. We next defined a weighted distance (WDC-C) by normalizing distances between the centers of the estimated partial sizes of tumors, and their potentially corresponding elevated metabolomic profile regions, with the number of involved voxels.

In addition to the above, our assumptions and approximations also extended to other confounding factors, such as magnetic field inhomogeneity and the possibility of specimen shrinkage during histology (an average reduction of 27.3 ± 7.7% was found in this study). It would thus be unlikely for any locations of voxels with suspicious metabolomic profiles of prostate cancer and any malignant regions identified by histology to overlap precisely. However, a corresponding relation between the two approaches (metabolomic profiles versus histology) was illustrated by the positive linear correlation between tumor size, as determined by histology, and the value of MIp (Fig. 4A). This apparently significant linear relation only existed among T2 tumors, probably because the current MIp was heavily skewed toward T2 tumors, as previously discussed. An inverse linear correlation was also found between the histologically and metabolomically identified regions by comparing the average intensities of metabolomic profiles for each voxel in the suspicious regions and the WDC-C (Fig. 4B). This inverse linear correlation supports our expectation that the closer the two centers, the higher the per-voxel value of the metabolomic profile. Further, although our findings are preliminary, owing to the limited number of studied samples, the capacity of the MIp index to differentiate between prostate cancer–positive and prostate cancer–negative regions is significant (Fig. 4C), with overall accuracies of 93% estimated for all five tumors and 97% for four T2 tumors (Fig. 4D). However, it should be emphasized that these negative regions still possess elevated values of metabolomic profiles.

Fig. 4.

Correlations of metabolomic profiles with histology. (A) Sizes of T2-stage tumors correlate linearly and positively with values of the malignancy index (r2 = 0.975, P < 0.013). Cancer lesions shown in Fig. 3 are shown as ▪, other T2 tumors as •, and the T3 tumor as ♦. (B) A significant inverse linear relation exists between the average intensities of metabolomic profiles for voxels in the involved profile-elevated regions and the weighted distances (WDC-C) (r2 = 0.998, P < 0.001) for T2 tumors. This indicates that the closer the histologically and metabolomically represented voxel centers, the stronger the cancer profile signal in the metabolomic map. (C) The malignancy index provides a threshold indication of malignant potential for profile-elevated regions with and without histologically identifiable cancer (P < 0.008, for all tumors; P < 0.004, for T2 tumors); overall accuracies are presented by the ROC curves (D) for all tumors and for T2 tumors only. AUC, area under curve.


The hypothesis underlying the cancer metabolomics evaluated in our current study involves three interlocking concepts. First, mechanisms that permit the growth of malignant tumors are unlikely to affect only one or a few metabolites. Next, alterations occur in overall measurable metabolites (that is, metabolomic profiles) in the presence of tumor growth. Finally, metabolomic profiles are more sensitive than any single metabolite in identifying and characterizing pathological disease (2427).

All these concepts are based on the well-established view that cellular metabolic status at the time of tissue excision is preserved within intact tissues. We previously proposed the utility of prostate cancer metabolomics (19); the current study, which assesses this proposition, evaluated the potential of metabolomic imaging to detect sites of human prostate cancer.

By applying prostate cancer metabolomic information from intact tissue analyses to evaluations of removed whole prostates obtained with human whole-body MRI, the current study evaluated the clinical potential of metabolomics to direct prostate cancer biopsy. Metabolomic analysis of the sizes of suspicious regions and distances (WDC-C) allowed marking of areas and locations of biopsy interest. For instance, using the MIp index, cancer lesions with an overall size of more than twice the MRS voxel size (~18 mm2, used for this study) may be detected in the center of a metabolomically identified suspicious region within a perimeter of 5.7 ± 1.5 mm, with an overall accuracy of 97% (T2 only) or 93% (including T3). Our illustration is likely to overestimate accuracy, given the small sample sizes examined. It nevertheless points toward the utility of metabolomic mapping to an important and vexing clinical dilemma.

Of further note, the prostate cancer MRSI literature supports the use of the ratio of spectral regions for choline-to-creatine over citrate (CC/Ci) in evaluating clinical in vivo MRSI results (12, 13, 28). Using data from the present study, we analyzed the CC/Ci maps in the same manner as the metabolomic profile presented here. None of the CC/Ci results in the analyzed voxels were able to indicate tumor regions with statistical significance. This is likely because of our sample size, which was much smaller than the reported MRSI studies. However, these results serve to illustrate the strength of metabolomic profiling in diagnosing individual patients, a matter of greater clinical interest than retrospective group comparison.

The current study was limited by technical factors. The small number of samples seen in this ex vivo study of five removed prostates, analyzed on a human 7-T MR scanner, precluded their use alone in establishing and testing metabolomic profiles. Data from these samples were supplemented by data from previous tissue analyses obtained with a 14-T spectrometer so as to permit the calculation of profile values. In the future, as sufficient numbers of samples for generating and testing cancer profiles are studied with the 7-T scanner, this inherent limitation will be overcome. Future studies may also better address confounding technical factors, particularly for magnetic susceptibility complications arising from the tissue-air interface, by in vivo implementation of the demonstrated metabolomic imaging concept. Here, such complications led to the exclusions of true cancer-positive voxels. However, the capacity of in vivo MR spectra to include prostate edge voxels obviates this ex vivo concern. Further in vivo study may permit the inspection of the entire prostate. Finally, metabolomic profiles may be sensitive to histologically benign prostate regions adjacent to a histologically defined cancer region, thereby augmenting other radiological approaches for detecting malignancy with data including the edge of the prostate.

Challenges already experienced in adapting in vivo prostate MRS from 1.5- to 3-T scanners indicate that implementing in vivo prostate metabolomic imaging on a 7-T platform will not be straightforward (2934). Nevertheless, metabolomic imaging has the potential to detect lesions, guide biopsy, and eventually identify other conditions of malignancy, such as tumor aggressiveness; it can be applied to evaluate other human malignancies and can extend beyond a mapping of tissue metabolites to include other disease-sensitive radiological parameters.

Materials and Methods

This study of human prostates was approved by the Institutional Review Board at Massachusetts General Hospital in accordance with an assurance filed with and approved by the U.S. Department of Health and Human Services. Five prostates removed from prostatectomy surgeries of patients with biopsy-proven prostate cancer were transferred on ice to a Siemens 7-T 90-cm human MR scanner. MRI and MRS were conducted at room temperature, with a scanner body coil as transmitter and a circular 7-cm-diameter surface coil as receiver. For accurate MRS voxel placement and coregistration with the pathology, T2-weighted images 1 mm thick were collected with a turbo spin-echo sequence on ~30 slices, with field of view (FOV) of ~40 mm × 40 mm (depending on prostate size).

Two-dimensional multivoxel proton MRS was acquired on three different levels of the prostate. The volume of interest was selectively excited using echo time/repetition time of 30/1700 point-resolved spectroscopy (PRESS) with chemical shift–selective (CHESS) water suppression. The FOV of ~40 mm × 40 mm was partitioned into 16 × 16 phase-encoding steps. The individual voxel size was ~3 mm × 3 mm × 3 mm. A spectral width of 4.0 kHz, water suppression bandwidth of 130 Hz, and 16 acquisitions were used. Each experiment required ~23 min, a time frame within which future in vivo implementations can be accommodated. MRS data, processed with NMR Software Nuts (Acorn NMR Inc.), determined the intensities of metabolites.

After MRI-MRS measurements, the entire prostate was fixed, sliced transversely at about 4- to 5-mm intervals so as to produce seven to nine serial block sections, and paraffin-embedded. One whole-mounted histopathology slide was obtained from each block. All slides were hematoxylin and eosin–stained and evaluated by the study pathologist to identify cancer regions throughout the entire prostate.

The value of the metabolomic profile for each voxel was calculated with modified parameters determined from proton MRS of intact human prostate tissues (19). The point spread function of the multivoxel MRS is not an ideal impulse function, so the value of the metabolomic profile for each voxel may have included contributions from neighboring voxels. A deconvolution process was used by filtering profile values with a 3 × 3 two-dimensional finite impulse response filter of the following coefficients: 1/4 for the center voxel of interest, 1/8 for the adjacent four voxels sharing one edge with the center voxel, and 1/16 for the four diagonally adjacent voxels. For edge voxels, the values were adjusted to compensate for the lack of data outside the image. To visualize the processed metabolomic profiles, we scaled the image data to the full range of a 64-color map and overlaid them onto the anatomical image.

Statistical methods

After PCA, PC j for a hypothetical sample X can be written as follows:PCj,X=Aj(cj,1*p1,X+cj,2*p2,X+cj,3*p3,X++cj,i*pi,X)=AjΣicj,i*pi,X(2)wherein Aj is an arbitrary constant that determines the mean of the PC to be 0, p represents the standardized peak intensity of resonance i, and cj,i is the PCA coefficient (also known as the loading factor) for PCj and pi determined by the eigenvectors of the correlation matrix of p.

By contrast, after canonical analysis, the canonical score k involving selected PCs (L, M, …N) for sample X would be represented as follows:Bk(eL,kPCL+eM,kPCM+...+eN,kPCN)=BkΣjej,kPCj=BkΣjej,k(AjΣicj,ipi,X)=Bk(Σjej,kAjΣi(Σjej,kcj,i)pi,X=CkΣifi,kpi,X(3)in which Bk, functioning similarly as Aj for PC, is a constant from canonical analysis, Ck is an overall constant involving constants A and B, ej,k is the canonical coefficient for PCj and score k, and fi,k = Σj ej,k * cj,i is the “overall loading factor” that for pi combines coefficients from PCA and canonical analysis.

All statistical analyses were conducted with JMP 7.0 software (SAS Institute Inc.). All measures of statistical significance were tested with two-sided P values.


  • * These authors contributed equally to this work.

  • Citation: C.-L. Wu, K. W. Jordan, E. M. Ratai, J. Sheng, C. B. Adkins, E. M. DeFeo, B. G. Jenkins, L. Ying, W. S. McDougal, L. L. Cheng, Metabolomic Imaging for Human Prostate Cancer Detection. Sci. Transl. Med. 2, 16ra8 (2010).

References and Notes

  1. Acknowledgments: We thank K. Isselbacher for encouragement and guidance, J. Fordham for editorial assistance, and G. Dai and G. Wiggins for initial technical support. Funding: CA115746 (L.L.C.); CA095624 (L.L.C.); A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital; and Claire and John Bertucci Prostate Cancer Research Fund. Author contributions: C.-L.W.: experimental design, funding and analysis of histopathology, and manuscript review; K.W.J.: MR experiments (performance and design); E.M.R.: supervision of MR experiments; J.S.: imaging analysis and presentation; C.B.A. and E.M.D.: literature reviews and manuscript editing; B.G.J.: experimental design assistance for MR imaging; L.Y.: supervision of imaging analysis and presentation of data; W.S.M.: recruitment of subjects and manuscript review; L.L.C.: articulation of conceptual framework, project design, funding support, and manuscript drafting and review. Competing interests: L.L.C., C.-L.W., and B.G.J. have applied for a patent on the method described in this paper. The authors report no other competing interests.
View Abstract

Stay Connected to Science Translational Medicine

Navigate This Article