Genitourinary Imaging
Original Research
Diffusion-Weighted MRI in the Detection of Prostate Cancer: Meta-Analysis
Audio Available | 
Share
OBJECTIVE. The objective of our study was to estimate and compare the performance of diffusion-weighted imaging (DWI) with other MRI techniques including T2-weighted MRI for the detection of prostate cancer.
MATERIALS AND METHODS. Searches of the PubMed and Scopus electronic databases for the terms “prostate,” “cancer,” “diffusion-weighted imaging,” and “magnetic resonance imaging” using an end date of December 2010 were completed. All included studies had histopathologic correlation; 2 × 2 contingency data were constructed for each study. A Bayesian receiver operating characteristic (ROC) model was used across studies to determine sensitivity, specificity, and area under the full or partial ROC curve.
RESULTS. Nineteen articles consisting of a total of 5892 lesions were analyzed. Based on a 95% credible interval, DWI alone yielded a significantly better area under the ROC curve, sensitivity, and specificity (0.85, 0.69, 0.89, respectively) than T2-weighted imaging alone (0.75, 0.60, 0.76). Combined DWI and T2-weighted imaging (0.73, 0.70, 0.83) showed a similar area under the ROC curve but significantly better sensitivity and specificity than T2-weighted imaging alone. DWI and combined DWI and T2-weighted imaging yielded similar overall sensitivity, but DWI alone showed better overall specificity than combined DWI and T2-weighted imaging. At specificities of greater than 80%, combined DWI and T2-weighted imaging yielded a partial area under the ROC curve (0.138) similar to that of DWI alone (0.129) and was significantly better than the partial area under the ROC curve of T2-weighted imaging alone (0.070). DWI alone and combined DWI and T2-weighted imaging appear to be superior to dynamic contrast-enhanced imaging alone (area under the ROC curve, 0.79; sensitivity, 0.58; specificity, 0.82).
CONCLUSION. DWI appears to improve diagnostic performance and can be a useful adjunct to conventional anatomic imaging for identifying tumor foci in prostate cancer.
Keywords: diffusion-weighted imaging, MRI, prostate cancer
High-resolution T2-weighted imaging has been the mainstay of MRI in view of its high tissue contrast resolution and ability to aid staging, including assessing extraprostatic extension of prostate cancer [1, 2]. In a meta-analysis by Sonnad et al. [3], T2-weighted imaging with T1-weighted imaging showed a maximum joint sensitivity and specificity rate of 74% for staging of prostate cancer. Engelbrecht et al. [4] reported similar findings, with a joint sensitivity and specificity rate of 71%. Hence, there exists a need for better ways to diagnose prostate cancer by MRI, particularly in cases in which the prostate-specific antigen level is rising and random ultrasound-guided needle biopsy (transrectal ultrasound [TRUS]) results are negative.
The introduction of nonsurgical local ablative techniques such as high-intensity focused ultrasound and cryotherapy [5] and advances in radiotherapy (e.g., brachytherapy, intensity-modulated radiotherapy, and proton radiotherapy) theoretically allow escalation of radiation dose to the dominant intraprostatic lesions without increasing dose to the surrounding normal tissues [6–8]. This capability further increases the need for accurate delineation of prostate cancer foci by noninvasive techniques.
However, T2-weighted imaging can be limited in distinguishing between tumor and benign disease such as inflammatory tissue. This shortcoming has led to substantial interest in MRI techniques—such as diffusion-weighted imaging (DWI), MR spectroscopy (MRS), and dynamic contrast-enhanced MRI (DCE-MRI)—that can be performed concurrently with anatomic imaging (primarily T2-weighted imaging) by MRI to improve detection of tumor foci in the local assessment of prostate cancer.
Compared with other MRI techniques, DWI has the advantages of a short acquisition time, no need for IV contrast material, and low technical demand for image postprocessing. DWI measures restriction of water diffusion in biologic tissues corresponding to properties such as cellular density, membrane permeability, and space between cells [9]. For example, the luminal space in benign human prostate tissue has been reported to average several hundreds of microns wide, whereas in prostate cancer water molecules diffuse over tens of microns [10]. This difference in water diffusivity may make it possible for DWI to differentiate malignant from benign prostatic tissues.
Typically, an ultrafast echo-planar T2-weighted sequence is used for DWI; this sequence can be performed with an endorectal coil only, endorectal and external phasedarray coils, or an external phased-array coil only. DW images are postprocessed to obtain apparent diffusion coefficients (ADCs). In clinical practice, this information is represented in image form as an ADC map. Analysis of ADC maps has the advantages of eliminating the effect of T2 shine-through that can be seen on DW images and of objectively measuring tissue diffusivity [11].
A growing number of clinical studies in recent years have evaluated the utility of DWI either in combination with or in comparison with other MRI techniques for the detection of prostate cancer, and these studies have reported various sensitivities and specificities. However, a meta-analysis of these clinical studies has not yet been performed and, given its widespread availability with most modern MRI scanners, is appropriate. The purpose of this meta-analysis was to estimate and compare the performance of DWI with other MRI techniques, including T2-weighted imaging, for the detection of tumor foci in the local assessment of prostate cancer.
| Materials and Methods |   |
|---|
This meta-analysis was written using the preferred reporting items for systematic reviews and meta-analyses statement as a reference [12, 13].
A comprehensive systematic review of the literature up to December 2010 was completed. The PubMed and Scopus electronic databases were searched. The literature search was limited to English-language published studies and human subjects. The following medical subject heading terms and keywords were used in the search: “prostate,” “cancer,” “diffusion-weighted imaging,” and “magnetic resonance imaging.” The abstracts of all relevant articles published before December 2010 available in the databases were reviewed.
The inclusion criteria for articles were as follows: patients had histologically proven prostate cancer; the diagnostic tests included DWI with or without other methods of imaging such as T2-weighted imaging, MRS, or DCE-MRI; the reference standard was histologic diagnosis; and sufficient data were reported to allow construction of 2 × 2 contingency tables. Only studies that used histopathologic results to directly reference independently derived DWI findings were included in our study; those in which the DWI findings were retrospectively obtained from prefabricated region-of-interest (ROI) maps using data from either histopathologic results or T2-weighted imaging were excluded. The study design inclusion criteria were broad and included retrospective and prospective studies. Articles were excluded if they were editorials, commentaries, or case reports.
The abstracts of 150 published studies from the PubMed and 244 published studies from the Scopus electronic databases were retrieved using the search terms. The combined records identified 105 English-language scientific published studies after removal of duplicates. Of these studies, 26 articles were excluded: studies performed on extraprostatic organs (nine published studies), studies of non-DWI techniques (11 published studies), studies of healthy subjects (five published studies), and an ex vivo study (one published study). Full-text evaluation of the remaining 79 published studies was performed. Among these studies, 60 were excluded for the following reasons: DWI data were retrospectively obtained using T2-weighted imaging, DCE-MRI, or histologic maps as a guide to draw DWI ROIs (41 published studies); there were insufficient data to construct 2 × 2 tables (15 published studies); TRUS-guided biopsies were based on MRI data (two published studies); local tumor stage rather than tumor location was used as the endpoint (one published study); and raters were not blinded to the histologic results (one published study).
Nineteen published studies [14–32] were included in the meta-analysis, and each reported one or more imaging sequences to classify prostate lesions as benign or malignant on the basis of imaging results. Data from the 19 selected articles were extracted for information on study design, year of publication, number of patients, number of lesions studied (total numbers and number of lesions positive for prostate cancer), number of raters, reference standard (whole-mount or step-section histopathology, TRUS-guided biopsy), imaging techniques (DWI, T2-weighted imaging, DCE-MRI, MRS), field strength, type of coil (torso surface phased-array, endorectal, body coil), imaging parameters (b value), and criteria for diagnosis (e.g., hyperintensity on DW images, hypointensity on ADC maps, ADC cutoff values).
For all 19 studies evaluated, the number of true-positive, false-positive, true-negative, and false-negative lesions for each imaging technique had to be available for analysis. The area under the receiver operating characteristic (ROC) curve, sensitivity, specificity, and accuracy data were recorded separately. All data for which the 2 × 2 tables could be constructed were used for analysis. We did not exclude any part of the studies included in the meta-analysis. Thereby, a total of 38 raters and 5892 lesions were included in the meta-analysis. The recorded variables for all the included studies are presented in Table 1.
Because most of the studies included in our analysis reported only sensitivity and specificity values for the modalities that they evaluated, we fitted a Bayesian ROC model to the sensitivity and specificity values reported in the studies to compare sensitivity, specificity, and area under the ROC curve across studies and modalities [33]. We assumed a binormal model for the generation of the classification of lesions. In this model, a continuous latent trait is associated with the manifestation of “disease” in each image. Because this model was used to construct ROC curves based only on published sensitivity and specificity values, it was necessary to assume homogeneity of the binormal model across studies. To better evaluate the use of DWI as an imaging technique, a partial area under the ROC curve for a specificity of more than 80% was performed because that degree of specificity desirable in clinical practice— specifically, for patients with prostate cancer—to avoid unnecessary treatment.
We assumed that the latent outcome variable associated with normal lesions was drawn from a standard normal distribution, whereas the latent outcome variable associated with malignant lesions was assumed to follow a normal distribution with a mean μd and an SD. We further assumed that the thresholds used in each study for each threshold to determine positive and negative findings were drawn from a normal distribution with a mean of μthreshold and an SD of SDthreshold.
Using a Metropolis-Hasting algorithm to perform Markov chain Monte Carlo sampling, we obtained the posterior distribution of the parameters specified earlier. This model allowed us to compute the posterior distribution on the total area under the ROC curve and the partial area under the ROC curve (i.e., cutoff point based on specificity) for each image modality across the population of studies considered. Comparisons between modalities were based on 95% credible intervals for the area under the ROC curve and partial area under the ROC curve values.
TABLE 1: ticles Selected for Meta-Analysis and the Various Data Recorded Ar
TABLE 2: Summary of Studies Reviewed by Modality
Among the various studies, only four imaging techniques could be successfully analyzed: six studies of DCE-MRI alone; 10, DWI alone; 14, T2-weighted imaging alone; and 11, joint T2-weighted imaging and DWI (henceforth referred to as “combined DWI and T2-weighted imaging”) (Table 2). Four studies reported results for each of the multiple raters. Inexperienced raters were identified by reviewing the original articles, and a sensitivity analysis was performed to assess the effect of inexperienced raters (three raters in total) on the performance of combined DWI and T2-weighted imaging.
An analysis of MRS alone or in combination with DCE, DWI, or T2-weighted imaging could not be performed because of the small numbers of published studies and sample populations (Table 2). Given the relatively small amount of data, further analysis of the selected studies regarding the use of DWI for the detection of disease in the central zone versus the peripheral zone, the optimal b value, the use of DW images versus ADC maps, and the use of 3-T scanners versus 1.5-T scanners as factors influencing lesion detection could not be made.
| Results | |
|---|
Figure 1 shows the estimated area under the ROC curve for each method. Table 3 summarizes the p distributions for all the parameters.
Based on a posterior 95% credible interval, the area under the ROC curve for T2-weighted imaging alone is significantly less than that for DWI alone but is not significantly different from that of combined DWI and T2-weighted imaging (T2-weighted imaging, 0.71–0.77; DWI, 0.82–0.87; combined DWI and T2-weighted imaging, 0.71–0.75). The overall sensitivity, specificity, and partial area under the ROC curve (specificities > 80%) of DWI alone and of combined DWI and T2-weighted imaging are superior to T2-weighted imaging alone. The sensitivity of T2-weighted imaging was 0.57–0.62; DWI, 0.67–0.72; and combined DWI and T2-weighted imaging, 0.69–0.72. The specificity of T2-weighted imaging was 0.74–0.78; DWI, 0.87–0.90; and combined DWI and T2-weighted imaging, 0.80–0.85. The partial area under the ROC curve of T2-weighted imaging was 0.06–0.09; DWI, 0.121–0.136; and combined DWI and T2-weighted imaging, 0.134–0.140.
At an extremely low specificity (i.e., the right part of the ROC graphs), the unusual shapes of the ROC curves of combined DWI and T2 for all raters and for experienced raters only are due to the fact that the 80% sensitivity was relatively constant and little data were available to very low specificity; therefore, the behavior at extreme values was uncertain.
The overall area under the ROC curve of combined DWI and T2-weighted imaging is significantly lower than that of DWI alone. However, when comparing at specificities of more than 80%, combined DWI and T2-weighted imaging yielded a partial area under the ROC curve similar to that of DWI alone.
Based on the studies, the area under the ROC curve (0.75–0.83) and overall sensitivity (0.53–0.62) of DCE-MRI were not significantly different from those of T2-weighted imaging, even though the specificity of DCE imaging (0.80–0.85) is increased. In contrast, DCE-MRI was less sensitive than DWI alone or combined DWI and T2-weighted imaging and was less specific than DWI alone; however, the specificity of DCE-MRI was not significantly different from that of combined DWI and T2-weighted imaging. The partial area under the ROC curve of DCE-MRI (0.07–0.09) is not significantly different from the partial area under the ROC curve of T2-weighted imaging but is significantly less than that of DWI and of combined DWI and T2-weighted imaging.
| Discussion | |
|---|
DWI for imaging the prostate is common-place in most institutions. An understanding of the utility of this sequence is important because DWI is being used to affect clinical care. A number of clinical studies have evaluated the utility of DWI alone or in combination with other sequences for the assessment of prostate cancers, and most of these scientific articles were published over the past 5 years. Hence, it is both important and appropriate to validate the efficacy of DWI in the detection of prostate cancer.
 View larger version (41K) | Fig. 1 —Estimated receiver operating characteristic (ROC) curves for imaging techniques in meta-analysis. Area under ROC curve was 0.75 for T2-weighted imaging, 0.79 for dynamic contrast-enhanced MRI (DCE-MRI), 0.85 for diffusion-weighted imaging (DWI), and 0.73 for combined DWI and T2-weighted imaging. |
Our meta-analysis revealed several important findings that pertain to the clinical use of DWI for the detection of prostate cancer. First, DWI alone yields a significantly better area under the ROC curve than T2-weighted imaging alone. This finding is in line with the results of published studies at both 1.5 T [6, 19, 21, 26, 34, 35] and 3 T [14, 20, 36] that show increased sensitivity and increased specificity of DWI. This better performance appears to represent an inherent advantage of DWI over conventional anatomic imaging alone, which is consistent with DWI assessing parameters different from T2-weighted imaging.
Second, at a defined specificity of greater than 80% (partial area under the ROC curve), combined DWI and T2-weighted imaging is superior to T2-weighted imaging alone and DCE-MRI alone (Table 3) but is similar to DWI alone. These findings suggest that to achieve high specificity, DWI should be added as an adjunct to T2-weighted imaging. Currently T2-weighted imaging is the principal sequence used for the evaluation of extraprostatic involvement in clinical practice.
TABLE 3: Overall and Random Sensitivity and Specificity Rates and Total and Partial Area Under the Receiver Operating Characteristic (ROC) Curve for Studies Grouped by Modality
Third, the total area under the ROC curve of combined DWI and T2-weighted imaging appears to be worse than that of DWI alone. A review of the published studies indicates that only two studies directly compared DWI and combined DWI and T2-weighted imaging [19, 24]. In both studies, the sensitivity of combining DWI with T2-weighted imaging was higher than that of DWI, even though the specificity rates did not improve. More studies performing head-to-head experiments like these are needed.
Because the data extraction for this meta-analysis was focused on DWI, the results assessing T2-weighted imaging and DCE-MRI are not fully inclusive and need to be considered in relation to the results of DWI for a meaningful analysis. On the basis of our meta-analysis, DCE-MRI alone appears to show specificity similar to combined DWI and T2-weighted imaging but is less specific than DWI alone. It is also less sensitive than both DWI alone and combined DWI and T2-weighted imaging. These findings suggest that DWI outperforms DCE-MRI and raises the question of whether adding more MR techniques, such as DCE imaging, to multiparametric imaging that already includes DWI and T2-weighted imaging increases diagnostic performance.
Technical parameters—Because the number of published studies that met the inclusion criteria was relatively small, there were insufficient data to compare several important DWI parameters—namely, magnet field strength (1.5 vs 3 T), type of coil (endorectal vs surface phased-array), and b value. As many as nine of 19 studies were performed using a surface phased-array coil on a 1.5-T scanner, but these conditions would be considered suboptimal to the use of endorectal coil imaging [37–39]. Of the 19 studies included in the meta-analysis, five studies used endorectal coil for imaging on 1.5-T scanners [19, 23, 24, 27, 40]; another five studies used surface phased-array coils for imaging on 3-T scanners [14, 17, 20, 25, 32].
Although the magnet field strength (1.5 vs 3 T) and type of coil are subject to the resources of each institution, the b value parameter is not. Even so, there has been no consensus about the optimal b value to date, and the b value used varies greatly among the studies included in our meta-analysis (Table 1). Although higher b values have been shown to enable better characterization of prostate cancer and treatment response in experimental mouse models [41], they also lead to increased motion and susceptibility artifacts as well as a decreased signal-to-noise ratio [42]. Studies directly comparing the results of DWI using different technical parameters are needed.
Methods of interpretation—We excluded studies that retrospectively analyzed the ADC value of lesions using ROIs drawn from other diagnostic tests (e.g., T2-weighted imaging, DCE-MRI, or histologic maps) because these methods could skew the data and do not allow proper evaluation of DWI as a diagnostic tool. Although using this selection process meant that we had to exclude 60 of 79 studies that were initially considered for full-text evaluation, omitting those studies from the meta-analysis allowed us to better analyze the true clinical value of DWI as an imaging test for prostate cancer detection. The selection process was thus designed to remove bias and create a more robust dataset.
In terms of lesion diagnosis, more than half of the studies used ADC maps for interpretation (12 of 19) either alone or in combination with DW images, whereas five studies relied on DW images alone. In our experience with prostate DWI, ADC maps can be more easily interpreted than DW images. On DWI, both tumor and normal tissue can be T2 hyperintense because of T2 shine-through effects, whereas T2 shine-through effects are removed from ADC maps; hence, on ADC maps, tumor will appear hypointense compared with normal prostatic tissue.
A limitation in the evaluation of the published studies pertains to combined DWI and T2-weighted imaging: In most studies, it was not explicitly stated what constituted a positive finding—that is, if DWI detects disease in areas of normal T2 signal or if DWI increases diagnostic confidence in areas with low T2 signal. The increased sensitivity of DWI alone compared with T2-weighted imaging alone suggests that DWI does detect disease in areas of T2 signal considered normal when read in isolation. It is likely that when DWI and T2-weighted imaging are read together, diagnostic confidence can also be increased. The results of this meta-analysis suggest that DWI alone is superior to combined T2-weighted imaging and DWI. However, because only one study [24] directly compared between them, we were not able to come to such a conclusion reasonably.
Reference standard for histologic assessment—In 12 of the 19 studies included in this meta-analysis, only TRUS-guided biopsy results were used to correlate with imaging findings (Table 1). Needle biopsies can result in sampling error and false-negatives; if taken to be the reference standard, needle biopsy results may underestimate the sensitivities of the imaging modalities. Furthermore, a per-nodule analysis, which requires direct histopathologic correlation in all cases with a step-section or whole-mount technique, is the optimal approach. This analysis was performed in seven of the studies in our meta-analysis, including one study that used either TRUS or step-section histopathology to correlate with imaging. However, this approach can be both cost- and labor-intensive; hence, routine application of these techniques may not be possible in many studies. Indeed, if we had restricted our meta-analysis to these gold standards, only six studies would have qualified, yielding too few studies for a meta-analysis.
The studies that used step-section or whole-mount technique as the reference standard revealed findings similar to the overall findings in our meta-analysis: First, DWI is more specific but less sensitive than combined DWI and T2-weighted imaging [19]; second, combined DWI and T2-weighted imaging is more sensitive and more specific than T2-weighted imaging alone [15, 19, 21, 24, 28, 29]; and, third, DWI is more specific than DCE-MRI [27], whereas combined DWI and T2-weighted imaging is more sensitive and more specific than combined DCE-MRI and T2-weighted imaging [28].
Study endpoints—In this meta-analysis, the outcomes considered were detection of tumor foci; evaluating the use of DWI for assessing extracapsular extension was not an aim of most studies. Traditionally, T2-weighted imaging has been considered to be the optimal technique for the anatomic delineation of the local extent of disease. Given the lower spatial resolution of DWI compared with T2-weighted imaging, DWI may theoretically have poorer performance in the assessment for extracapsular extension. In comparison, the results of studies in the literature suggest that invasion of other structures such as the seminal vesicles [36, 43] and bladder [44] can be detected better on combined DWI and T2-weighted imaging than on T2-weighted imaging alone; however, this question requires further study. The results of this meta-analysis cannot be used to suggest that DWI is sufficient to replace T2-weighted imaging and can be read in isolation for prostate cancer staging assessment.
Comparison of DWI, DCE-MRI, and MRS— As multiparametric MRI assessment of the prostate becomes more widely available, comparisons of DWI against or combined with other MRI techniques, such as DCE-MRI and MRS, need to be studied to address the issue of MRI protocol optimization.
Based on our meta-analysis, DCE-MRI appears to be inferior to DWI—alone or in combination with T2-weighted imaging—in terms of overall sensitivity and partial area under the ROC curve. However, the data did not allow additional comparisons, such as between MRS and combined T2-weighted imaging and DCE-MRI. To achieve adequate comparison of techniques, it is necessary to include studies that evaluate DCE-MRI or MRS without DWI; however, this comparison is beyond the scope of the current review. The role of multiparametric MRI could not be fully addressed in this study because of the small number of studies combining the various imaging techniques. It is possible that more may not be better because there may not be significant benefits in test characteristics associated with a three-parameter analysis (i.e., DWI, DCE, and T2-weighted imaging) over a two-parameter analysis (combined DWI and T2-weighted imaging), as suggested in a recent study by Vilanova et al. [45], and because of additional study time and cost.
Potentially important uses of DWI, such as improving diagnosis of prostate cancer in the central zone and in posttreatment states, could not be evaluated in this meta-analysis because of limited data. The results of a few individual studies suggest that there may be a role for DWI. Shimofusa et al. [21] found that DWI detected tumor in the central zone in five of eight patients (63%) compared with tumor being detected using T2-weighted imaging in one of eight patients (13%). In patients who underwent high-intensity focused ultrasound for the detection of prostate cancer, combined DWI and T2-weighted imaging has been shown to be more specific than DCE-MRI (74–78% vs 63–68%, respectively) in detecting posttreatment tumor progression, although sensitivity may be lower (63–70% vs 80–87%) [14]. Given the relatively longer survival times of patients with prostate cancer as compared with other common adult cancers and the acceptable practice of watchful waiting of patients with early prostate cancer and a life expectancy of more than 10 years, there is a constant need to avoid overdiagnosis of prostate cancer. Currently, how the results of DWI impact the overall management of prostate cancer—for example, in terms of patient morbidity and survival—remains unanswered.
In conclusion, the results of this meta-analysis suggest that DWI is a useful adjunct to conventional anatomic imaging with T2-weighted imaging for detecting cancer within the prostate gland.
| References | |
|---|
