Accelerated T₂ Mapping for Characterization of Prostate Cancer

Abstract

Prostate T₂ mapping was performed in 34 consecutive patients using an accelerated multi-echo spin-echo sequence with four-fold k-space undersampling leading to a net acceleration factor of 3.3 on a 3T scanner. The mean T₂ values from the accelerated and conventional, unaccelerated sequences demonstrated a very high correlation (r = 0.99). Different prostate segments demonstrated similarly good interscan reproducibility (p = NS) with slightly larger difference at base: 2.0 ± 1.6 % for left base and 2.1 ± 1.1 % for right base. In patients with subsequent targeted biopsy, T₂ values of histologically proven malignant tumor areas were significantly lower than the suspicious looking but non-malignant lesions (p<0.05) and normal areas (p<0.001): 100 ± 10 ms for malignant tumors, 114 ± 23 ms for suspicious lesions and 149 ± 32 ms for normal tissues. The proposed method can provide an effective approach for accelerated T₂ quantification for prostate patients.

INTRODUCTION

Prostate cancer is the most common non-cutaneous tumor, and the second most common cause of cancer related death in American men (1). Magnetic resonance imaging (MRI) is currently regarded as the best, non-invasive modality for localization and staging of prostate cancer due to its superior soft tissue contrast and high resolution. T₂-weighted turbo spin-echo MRI is the mainstay sequence for depicting prostatic zonal anatomy. Approximately 70% of prostate cancers are found in the peripheral zone (PZ) of the prostate. The primary MR signature of prostate cancer in the PZ is a hypointense region within the normally hyperintense background from healthy glandular tissues on T₂ weighted images (2). However, T₂-weighted imaging alone is reported to have a wide range of sensitivity and specificity of 22 – 85% and 50 – 99%, respectively (2,3), motivating the exploration of additional MR-based techniques to improve the localization and staging of prostate cancer including MR-spectroscopic imaging, diffusion weighted imaging (DWI), and dynamic contrast enhanced (DCE) imaging. Generally overlooked is the potential value of quantitative T₂ mapping. Measurements of tissue T₂ values may prove of value in better discriminating cancer from healthy tissues(4,5) and will allow investigators to perform longitudinal studies and inter-scanner and intra-scanner comparisons. Moreover, T₂-weighted images have to be digitized in order to be fed into automated decision support systems, while quantitative T₂ maps can be incorporated directly.

The standard T₂ mapping involves acquiring spin-echo images at increasing echo times (TE) but is associated with long scan durations making this method impractical for most clinical applications. In previous prostate studies, sampling of the relaxation curve was limited to four to six points to reduce the scan time (4–9). Roebuck et al used a Carr-Purcell-Meiboom-Gill (CPMG) sequence to sample the T₂ decay with a series of equally spaced refocusing pulses in each repetition time to shorten the scan time (10). While it is more time-efficient than successive acquisition of single-echo spin-echo data, the multi-echo approach still results in rather long scan times because of the long repetition time for each phase encoding step. As a consequence, these measurements are prone to motion artifacts, especially if large volume coverage of the target anatomy is required. Furthermore, the associated specific absorption rate (SAR) is high due to the frequent use of 180° RF pulse series.

Recently, a new reconstruction algorithm combined with a dedicated k-space undersampling pattern was developed to reduce the number of phase-encoding steps in a multi-echo spin-echo measurement without loss of spatial resolution and dynamic range (11). In the proposed approach, the k-space data at each echo time were undersampled and a reconstruction algorithm that exploited the temporal correlation of the MR signal in k-space was used to reconstruct alias-free images. With this method, alias-free brain T₂ maps were acquired with four-fold k-space undersampling resulting in a net acceleration factor of 3.4 in healthy volunteers (12). The purpose of the present study was to characterize T₂ values of tumorous and normal prostate tissues using this accelerated T₂ mapping technique. Lesions that were suspicious for malignant tumor on MRI were confirmed with a targeted prostate biopsy system that fused real-time transrectal ultrasound (TRUS) images with previously acquired MR images (13,14).

MATERIALS AND METHODS

Study Design and Patient Population

This prospective, single institution study was approved by the local institutional review board (IRB) and was compliant with the Health Insurance Portability and Accountability Act (HIPAA); informed consent was obtained from each patient. A total of thirty-four consecutive patients were enrolled in the study. All patients underwent prostate MRI for localization and staging of prostate cancer. Because the present study focused on characterization of malignancy in the PZ zone, eighteen patients with suspicious lesions in the PZ zone were selected for a subsequent targeted biopsy by fusing the previously acquired MR images with TRUS imaging. The mean age of the patients with biopsy confirmed malignant tumor was 59 years and the mean prostate-specific antigen (PSA) level was 6.5 ng/dl (Table 1). The mean age for the patients with negative biopsy was 61 years and the mean PSA was 5.3 ng/dl.

Table 1.

Patients with MRI/TRUS fusion targeted prostate biopsy (median values are presented in parentheses).

	Number of patients	Number of biopsy cores	Mean Age(years)	Mean PSA(ng/dl)	Mean Tumor Gleason Score
Patients with Positive biopsy	7	13	59 (61)	6.5 (6.2)	7 (7)
Patients with Negative biopsy	11	18	61 (60)	5.3 (4.9)	n/a

Accelerated T₂ Mapping Technique

In a multi-echo spin-echo sequence with constant echo-spacing ΔTE, the k-space data S_i(k) at k-space positions k and consecutive echo times TE_i are linearly correlated by means of a convolution kernel E(k) (12),

$${\mathit{S}}_{i+1}(\mathit{k})={\mathit{S}}_i(\mathit{k})\otimes \mathit{E}(\mathit{k})\text{ with }\mathit{E}(\mathit{k})={\displaystyle \underset{x}{\int }{\mathit{e}}^{-\frac{\Delta TE}{T_2(x)}}{\mathit{e}}^{-i2\pi k\cdot x}\mathit{d}\mathit{x}}$$ [1]

provided the transverse magnetization follows approximately a mono-exponential decay. The convolution kernel E(k) has a well localized peak at k = 0 that diminishes quickly towards higher frequencies. This indicates that data samples that are closely neighboring in k-t-space are highly correlated, with the degree of correlation rapidly diminishing with the distance between the samples.

This high intrinsic redundancy in multi-echo spin-echo acquisitions is utilized to reduce the amount of data that is needed to fill k-space. The k-space undersampling scheme is illustrated in Figure 1 for an undersampling factor of R = 4 (11,12). Only one in every R phase-encoding steps (black lines) is acquired at each echo time skipping the rest of the phase-encoding steps (blue lines). The position of the measured k-space line is shifted from one echo time to the next to achieve a better coverage of k-space over time. Therefore, each k-space line is acquired once after R echoes. A number of blocks consisting each of R consecutive k-space lines are acquired in the centre of k-space without undersampling (red lines) for calibration purposes.

Figure 1.

Acquisition scheme for an undersampling factor R = 4. Black lines represent the acquired k-space samples. Blue lines represent the skipped sampling positions corresponding to Nyquist sampling. Red lines represent the samples acquired for autocalibration.

To avoid fold-over artifacts due to undersampling, the missing k-space samples are first estimated by exploiting the linear correlation between the k-space samples at consecutive echo times. Each missing sample S_i(k) is estimated by a linear combination of its neighbors in the k-t-space (11,12), according to:

$${\widehat{\mathit{S}}}_i(\mathit{k})={\displaystyle \sum _{\tau =-n_{\tau }}^{n_{\tau }}{\displaystyle \sum _{\varsigma =-n_{\zeta }}^{n_{\zeta }}{w}_{\tau }(\zeta )}}{\mathit{S}}_{i+\tau }(\mathit{k}+\zeta ),$$ [2]

where n_τ and n_ζ define the size of the neighborhood in the echo time and in k-space dimension respectively, and w denotes the coefficients of the linear combination. The k-space index k is either of dimension 2 for 2D acquisitions, or of dimension 3 for 3D acquisitions. Because the reconstruction kernel w is translation-invariant in the k-t-space, the coefficients w in Eq. [2] can be estimated from a subset of data that are acquired with Nyquist sampling for calibration purpose. The reconstruction coefficients w are calculated on the basis of a least-squares fit from the calibration samples:

$${\displaystyle \widehat{\mathbf{w}}}=\arg \underset{{\displaystyle \mathrm{w}}}{\min }{\displaystyle \sum _{i,k}{\Vert {\mathit{S}}_i(\mathit{k})-{\displaystyle \sum _{\tau ,\zeta }{\mathit{w}}_{\tau }(\zeta ){\mathit{S}}_{i+\tau }(\mathit{k}+\zeta )}\Vert }^2}.$$ [3]

MR Imaging

Prostate MRI scans including triplane T₂-weighted, DWI, DCE, 3D spectroscopic imaging, and T₂ mapping were performed using the 16-channel anterior half of a 32-channel SENSE cardiac array (Invivo, Orlando, FL, USA) in combination with an endorectal coil (BPX 30, Medrad, Warrendale, PA, USA) on a 3.0 T whole-body scanner (Achieva 3.0T, Philips Healthcare, Best, the Netherlands) on all patients. Triplane T₂-weighted turbo spin-echo images were acquired with TR / TE = 8852 / 120 ms, Field of view (FOV) = 140 mm × 140 mm, 3 mm slice thickness. DCE MRI was acquired with TR / TE = 5.5 / 2.1 ms, FOV = 262 mm × 262 mm , 6 mm slice thickness, obtained during a single-dose bolus injection of 0.1 mmol / kg gadopentetate dimeglumine injected at 3cc/sec. MR spectroscopy was acquired with 3D CSI PRESS, TR / TE = 980 / 100 ms, FOV = 72 mm × 72 mm, 6 mm slice thickness. Prostate T₂ maps were acquired with a multi-echo spin-echo sequence with an undersampling factor of R = 4 and 16 calibration lines. The net acceleration factor with respect to the conventional multi-echo spin-echo sequence was 3.3. The following parameters were applied for imaging acquisition: TR = 2200 ms, FOV = 280 mm × 280 mm, data matrix = 256 × 256, 3 mm slice thickness, 10 axial slices. Sixteen echoes were acquired with TE = 32 ms, 48 ms, 64 ms, …, up to 272 ms. The total scan time for T₂ mapping was 5 min 48 sec. In eight patients without targeted biopsy, another T₂ map with the same parameters was acquired about 30 minutes after the first T₂ map scan without repositioning the patients to investigate interscan reproducibility. In another eight patients without targeted biopsy, a single axial slice T₂ map was acquired with the same parameters with the conventional, unaccelearated sequence for reference.

Tumor “Ground Truth” Labels with Targeted Biopsy

In eighteen patients, target lesions for MRI/TRUS fusion guided biopsies were identified by an experienced radiologist based on multiparametric MR images (The T₂ maps were not used to define the target lesions). The criteria for positive lesions was defined in prior studies (15,16) as circumscribed, round/ellipsoid low-signal intensity areas on T₂-weighted images and/or low-signal intensity areas on ADC maps, and/or areas with a choline:citrate ratio of 3 or more standard deviations above the normal value on spectroscopic images, and/or nodular masslike hot spots on K^trans, K_ep maps from DCE MRI. During the biopsy, previously acquired MR images were fused to real-time TRUS, thereby the needle was inserted into the suspicious regions identified on the MRI using ultrasound guidance (13,14).

Data Analysis

Image reconstruction and T₂ mapping

The undersampled k-space data were transferred to a PC for data analysis with an in-house built software tool using C++. The missing data were reconstructed according to Eq. [2]. The convolution kernel matrix used for reconstruction of the undersampled images was chosen to be 3 × 3 × 9 (kx, ky and t) using the method described previously (11,12). T₂ maps were calculated pixel by pixel by fitting a monoexponential decay to the magnitude of the multi-echo spin-echo images by means of a non-linear least-squares algorithm. Only data above 10 times the standard deviation of noise were analyzed. The noise level was determined from the average of a rectangular area containing 20 × 20 pixels on the left top corner of the image at the first echo time for each slice.

Comparison with the conventional multi-echo spin-echo technique

To investigate the agreement of the accelerated T₂ mapping sequence with the conventional T₂ mapping sequence, mean T₂ values were computed in four ROIs drawn on the single slice from the conventional scan and identically placed ROIs drawn on the corresponding slice from the accelerated scan for each patient. Correlation between the mean T₂ values obtained with and without acceleration was assessed and Bland-Altman analysis was performed.

Interscan reproducibility

For the two T₂ maps acquired 30 minutes apart, six ROIs were drawn for each patient on six different prostate segments: right apex (RA), left apex (LA), right middle (RM), left middle (LM), right base (RB) and left base (LB). The reproducibility was calculated as the normalized value of the differences between the mean T₂ values from the two scans for each ROI.

T₂ statistics for malignant tumors

For patients with biopsy results positive for prostate cancer, tumor ROIs were drawn on the suspicious regions previously identified by the radiologist and confirmed with the targeted biopsy. Control ROIs were drawn on the contralateral side of the prostate if it was not suspicious looking on the multiparametric MR images according to the criteria defined above or on the same region on a neighboring slice if the contralateral region was suspicious. Comparisons of T₂ values between tumor and control areas were performed by paired Student t-test using Microsoft Excel.

RESULTS

T₂ maps were successfully obtained in all 34 patients. Reconstructed images from undersampled data exhibited no visible folding artifacts. Figure 2a and 2b illustrate the reconstructed images acquired at TE = 80 ms and TE = 160 ms from a patient (Yellow arrows indicate a suspicious hypointense area). The T₂ map calculated from the reconstructed multi-echo images is illustrated in Figure 2c with the suspicious lesion depicted as a dark spot. In Figure 3a, the accelerated T₂ map demonstrated a very good agreement with the reference, unaccelerated approach. A noticeable drop of SNR was observed in the conventional T₂ map (R =1) which was acquired with single slice acquisition which led to a shorter TR. The mean T₂ values obtained with R = 4 exhibited a strong linear correlation (r =0.99) with the mean T₂ values obtained with the conventional sequence (Figure 3b). The slope of the regression curve was very close to 1 with a small intercept of approximately 6 ms. The Bland-Altman plot confirmed the agreement between the conventional and accelerated measurements (Figure 3c). The mean difference between the two methods was 1 ± 7 ms. The interscan reproducibility of the accelerated T₂ mapping technique is illustrated in Figure 4. All six prostate segments demonstrated similarly good reproducibility (p = NS) with slightly larger difference at base: 2.0 ± 1.6 % for LB and 2.1 ± 1.1 % for RB.

Figure 2.

The reconstructed images at echo times 80 ms (a) and 160 ms (b) and the calculated T₂ map (c) from the reconstructed images from a patient (54 yrs, PSA = 3.6, Gleason 6 (3 +3)). Yellow arrows indicate a suspicious lesion.

Figure 3.

T2 maps (a) acquired from the accelerated approach (R = 4) and the conventional unaccelerated approach (R =1) from a patient(59 yrs, PSA = 11.1), regression plot (b) and Bland-Altman plot (c) of the mean T₂ values obtained with the accelerated T₂ mapping technique and the conventional T₂ mapping. The mean T₂ values obtained with R = 4 demonstrated very good correlations (r = 0.99) with the mean T₂ values obtained with the conventional sequence. The Bland-Altman plot confirmed the very good agreement between the regular and accelerated measurements.

Figure 4.

Interscan reproducibility of the fast T₂ mapping technique. All six segments demonstrated similarly good reproducibility (p = NS) with slightly larger difference at base: 2.0± 1.6 % for LB and 2.1 ± 1.1 % for RB.

Figure 5 illustrates a representative T₂ map from a 71-year-old male (a) and the corresponding lesions for fused MRI/TRUS guided targeted biopsy overlaid on T₂-weighted image (b). Two biopsy targets (red arrow heads) were identified based on multiparametric MRI (15,16). The left apical lesion was characterized by low signal intensity on T₂-weighted image, low signal intensity on ADC map, moderate hypervascularity on DCE MRI and the right apical lesion was positive on T₂-weighted image and ADC map only (data not shown). MR spectroscopy was not contributory to the diagnosis. Fused MR/TRUS guided biopsy targeting these two lesions revealed a malignant tumor with Gleason score = 7 (3 + 4) in the left lesion and a malignant tumor of Gleason score 6 (3 + 3) in the right lesion. These two lesions were conspicuous as hypointensities against the bright background on the T₂ map (Figure 5a). The T₂ values of the two lesions were 95 ± 7 ms for the right target and 97 ± 13 ms for the left target vs. 169 ± 17 ms for the right control and 272 ± 57 ms for the left control (Figure 5c).

Figure 5.

Representative T₂ map (a) and the corresponding lesion targets for fused MRI/TRUS guided biopsy overlaid on the corresponding T₂ weighted image (b). Fused MR/TRUS guided biopsy targeted for these two lesions revealed a malignant tumor of Gleason 7 (3 + 4) for the left lesion and a malignant tumor of Gleason 6 (3 + 3) for the right lesion. These two lesions were conspicuous on the T₂ map (a) with T₂ of 95 ± 7 ms for the right target and 97 ± 13 ms for the left target vs. 169 ± 17 ms for the right control and 272 ± 57 ms for the left control (c).

In 18 patients pre-biopsy MR imaging revealed 31 target lesions suspicious for prostate cancer. After fused MRI/TRUS guided targeted biopsy, 13 target lesions in 7 patients were histologically confirmed as prostate cancer (Table 1). The T₂ values of malignant tumors ranged from 85 ms to 124 ms and the T₂ of suspicious areas with negative biopsy findings fell in the range of 83 ms to 168 ms. The control areas had T₂ values all above 113 ms. Statistically, the T₂ values of malignant tumors were significantly lower than the control regions: 100 ± 10 ms for malignant tumors and 149 ± 32 ms for controls (p < 0.001) as shown in Figure 6. The T₂ values of suspicious regions (with negative biopsy) were lower than the controls (p<0.001): 114 ± 23 ms for suspicious regions vs. 153 ± 34 ms for controls. Compared to malignant tumor areas, the suspicious areas demonstrated higher T₂ values (p<0.05) while the control areas from both groups were statistically equivalent (p =NS).

Figure 6.

The T₂ values of malignant tumors were significantly lower than the control regions: 100 ± 10 ms for malignant tumors vs. 149 ± 32 ms for controls (p< 0.001). The T₂ values of suspicious regions (with negative biopsy) were also lower than the controls: 114 ± 23 ms for suspicious regions vs. 153 ± 34 ms for controls (p<0.001). Compared to malignant tumor areas, the suspicious areas demonstrated higher T₂ values (p<0.05) while the control areas from both groups were statistically equivalent (p =NS).

DISCUSSION

Conventional quantitative T₂ mapping using the multi-echo spin-echo sequence is associated with rather long scan times because of the long repetition time for each phase encoding step. In this study, T₂ values of prostate tissues were characterized with our new accelerated multi-echo spin-echo T₂ mapping technique in 34 patients. With k-t-reconstruction utilizing the temporal and spatial correlation of T₂ signal decay, fold-over-free images were reconstructed at each echo time providing a series of diagnostic images with variable T₂-weighting. As shown in Figure 2, the conspicuity of the suspicious lesion varied with the T₂ weighting as the lesion was better depicted on the later echoes. In a group of eight patients for which conventional multi-echo spin-echo data were also acquired, the prostate T₂ maps acquired with the proposed approach demonstrated excellent agreement with the conventional T₂ maps. The interscan reproducibility study performed in eight patients indicated that T₂ maps acquired with the accelerated sequence could be very well reproduced. The slightly worse reproduciblity at the base is likely due to the filling of the bladder over time. Our results demonstrate that this accelerated T₂ relaxometry can provide an effective approach to achieve T₂ quantification in prostate cancer patients.

The majority of the previous prostate T₂ measurements were performed on 1.5T. The reported T₂ values of malignant tissues ranged from 100 ms to 89 ms and T₂ values of normal tissues ranged from 193 ms to 112 ms (5,8,10,17). There has been only limited investigation of T₂ values of the prostate on 3T. Nevertheless, the T₂ values from the accelerated T₂ mapping technique were comparable to a recent report (18) : 100 ms vs. 109 ms for mean T₂ values of malignant tumors and 149 ms vs 142 ms for mean T₂ values of control tissues. We have also investigated suspicious but non-malignant areas which demonstrated higher T₂ (larger variations) than the malignant areas. Compared to suspicious areas, malignant tumors had lower T₂, therefore setting a threshold T₂ for suspicious lesions could improve the specificity of T₂-weighted imaging for cancer detection.

An undersampling factor of R = 4 used in the current study resulted in a net acceleration factor of 3.3. Due to scan time and SAR consideration, 10 slices were acquired with a scan time of approximately 6 minutes which was about the same as for T₂-weighted imaging. However, 10 slices (3 mm thickness) was not able to cover the whole prostate gland. The undersampling factor was typically selected by comparing the root-mean-square error (RMSE) of the reconstructed images with different kernel size and calibration lines. In this study, the undersampling factor was chosen based on our previous investigations which demonstrated a very good agreement between the brain T₂ maps acquired with an undersampling factor of R = 4 and those acquired with the conventional unaccelerated approach in volunteers (12). Increasing the undersampling factor can lead to further reduction in the scan time at the cost of lower SNR and higher residual fold-over artifacts on the reconstructed images. Since the reconstructed images in this study showed no visible folder-over artifacts and very good SNR, higher acceleration factors might be used for future scans which could achieve high resolution full-gland coverage T₂ mapping within scan time comparable to T₂-weighted imaging. To correlate MR findings with pathology is challenging for prostate. It is frequently difficult to obtain the “gold standard” prostatectomy histology because patients are on “watchful waiting” protocols or undergo radiation therapy. Targeted biopsy is an alternative method of validation that employs a stereotactic guidance. The biopsy core was taken from the target defined on MR images therefore providing validation for the ROIs drawn on T₂ maps (19).

It should be noted that the assumption of mono-exponential decay is used to derive an echo time invariant reconstruction operator that performs a convolution of the undersampled data. The reconstructed data themselves are not imposed to be mono-exponential. Indeed, previous works reported non-mono-exponential behaviors of the prostate T₂ relaxation (20). Echo time dependency associated with multi-exponential decays can be introduced into the reconstruction process by setting the temporal dimension of the convolution kernel equal to the number of echo times. Nevertheless, there are several limitations in our current study. First, the control areas were not biopsied. Nevertheless, the T₂ values of healthy tissues from this study agreed very well with a previous report (18) thus are unlikely to be false negative. Second, in the reproducibility study, patients were not actually repositioned between the two scans because it would have required two placements of the endorectal coil. Instead, we reperformed the acquisition with a 30 minutes interval. Finally, our study population and number of tumors analyzed are relatively small; larger patient populations will need to be studied to refine our results.

In conclusion, the accelerated T₂ relaxometry can provide an effective approach to speed up T₂ quantification with no apparent loss of lesion conspicuity. This method could potentially be incorporated into the standard clinical multiparametric MRI protocols providing a more quantitative approach to improve the detection and staging of prostate cancer.