Quantitative multiparametric MRI of ovarian cancer.

Ovarian cancer is the leading cause of death from gynecologic cancers among women in the United States ( 1 ). Five-year survival rates decrease dramatically from 90% in stage I to 30-40% in stage III and IV patients ( 2 ). The near absence of symptoms in early stage and the lack of effective screening tools both contribute to the poor overall survival. Several diagnostic modalities, including ultrasonography (US), computed tomography (CT), and MRI have been investigated as potential diagnostic and screening tools, and limited data demonstrates sufficient accuracy to predict malignancy either alone or as a part of screening strategy ( 3 - 14 ). Conventional contrast-enhanced MRI has an 84%-93% accuracy to differentiate between malignant and benign lesions ( 11 - 13 ). The interpretation of these images, however, is subjective and may vary depending on the experience of the reader ( 12 ). There are a number of objective physical parameters that can be measured with MRI, including relaxation rates, permeability and perfusion parameters derived from dynamic contrast enhanced (DCE) MRI, and diffusion-related parameters. Independently, several of these parameters have been found to be associated with ovarian malignancy: malignant masses have been shown to have shorter T 2 values ( 14 ), increased perfusion based on DCE-MRI parameters ( 15 , 16 ), and shorter water apparent diffusion coefficients (ADC) ( 17 ) than benign masses. In addition, a combination of DWI parameters, such as T 2 and b 1,000 signal intensity can differentiate benign from malignant ovarian masses ( 18 ). Improved imaging techniques could lead to development of screening modalities, which would increase the likelihood of detecting ovarian cancer at an early stage, and therefore improve survival. In addition, the development of techniques to apply to small volume disease could improve treatment monitoring for ovarian cancer patients. This study was designed to assess the feasibility of making multiple quantitative parametric measurements within a single study, and to identify which parameters can discriminate between benign and malignant masses. Additionally, we sought to determine whether parameter heterogeneity within a region of interest (ROI), represented by a combination of histogram characteristics, could provide additional information for classifying masses as benign or malignant beyond the mean parameter value. We therefore developed a quantitative multi-parametric MRI protocol, applied it in a pilot study of patients with ovarian masses, and compared our findings to surgical pathology. The aim of this study is to identify quantitative markers that can be used in studies for future development for the detection of early stage ovarian cancer and treatment monitoring.

Approval for this study was obtained from our Institutional Review Board (IRB). Thirtyseven subjects who were scheduled for surgical removal of an ovarian mass suspicious for malignancy by a gynecologic oncologist (based on imaging with either conventional MRI, CT, or US), had no contraindications for MRI were enrolled consecutively, and gave written consent. Subjects with a contraindication to receive gadolinium-DTPA or who had a glomerular filtration rate < 60 mL/min/1.73 m 2 were allowed to participate in the study without the administration of contrast or performance of DCE-MRI. All subjects were evaluated with quantitative MR scanning prior to surgical resection with histological verification. MRI scans were performed using a 3 Tesla MR scanner (TIM Trio, Siemens, Erlangen, Germany). The images were acquired with patients in the supine position and fitted with standard body matrix and spine array receive coils. Glucagon (0.5 mg intravenous) was given twice, once before scanning and again just prior to the DCE-MRI sequences to minimize small bowel motility in 12 patients (IRB approval of glucagon occurred mid-study). After localizer scans, anatomical imaging was performed using conventional multi-slice T 2 -weighted fast spin-echo images in the axial plane (parameters for all sequences provided in Table 1 ), and also the coronal and sagittal planes, followed by 3-dimensional T 1 -weighted fat-suppressed axial images covering the pelvis from the pubic symphysis to the sacral promontory. The anatomical images were then used to identify a target region that included solid regions of the mass if present. Three quantitative imaging series (detailed below) were then analyzed in the axial plane over the target region: T 2 mapping, ADC mapping, and DCE imaging. Reconstructions for all three quantitative methods were performed offline using Matlab (Mathworks, Natick, MA) to produce parameter maps in DICOM format for subsequent analyses. The T 2 mapping acquisition used a two-dimensional multi-echo spin-echo sequence (parameters in Table 1 ). The multi-echo image set was thresholded to eliminate noise pixels and then fit pixel-by-pixel to a monoexponential decay function to create a T 2 parameter map. Diffusion weighted images (DWI) were acquired using a single-shot fat suppressed echo-planar imaging sequence with three b-values (b=0, 100, 800 s/mm 2 ), three orthogonal diffusion traces, and four averages. All three b-value images were noise-masked, log-transformed, and then fit pixel-by-pixel with a linear model to produce parametric maps of the water ADC. Additional diffusion parameters were calculated using the theory of intravoxel incoherent motion (IVIM) proposed by Le Bihan et al. ( 19 ), which aims to separate true diffusion from perfusion effects. Following this method, the b=100 s/mm 2 and 800 s/mm 2 images were used to calculate maps of the true diffusion coefficient IVIM-D, and subsequently the perfusion fraction IVIM-F p . DCE acquisitions consisted of a series of forty 3D fat-suppressed gradient echo volumes, with a temporal resolution of 8.45 s/frame. Intravenous contrast injection (0.1 mmol/kg, 3 mL/s, 20 mL saline flush) was started after completion of four complete frames. The dynamic time series was noise masked and analyzed pixel-by-pixel using three previously established techniques. First, the normalized signal intensity was fit to a sigmoid function, as previously shown in ovarian cancer by Thomassin-Naggara et al. ( 16 ), to estimate the enhancement amplitude (EA), time of half rising T h , and maximal slope (MS) of the sigmoid curve. Secondly, the signal intensity was converted to a concentration by assuming a fixed pre-contrast T 1 value of 1500 ms, as suggested by Guo et al. ( 20 ). The first 60 seconds after contrast injection were integrated to calculate the initial area under the Gadolinium concentration curve (IAUC60). Finally, assuming a fixed, population averaged arterial input function ( 21 ), the Gd concentration curve was fit to the extended Tofts’ two-compartment model ( 22 ) to estimate the transfer constant K trans , the rate constant k ep , and the relative volume of the extravascular extracellular space, v e . In addition to these eleven quantitative parameters, normalized values for IAUC60, EA, and the axial T 2 -weighted anatomical image were produced by dividing by the mean value of an ROI placed adjacent skeletal muscle (piriformis when possible, gluteus maximus otherwise) to normalize for variation in cardiac output and system calibrations between subjects. The quantitative parameters assessed are summarized in Table 2 . A blinded radiologist (6 years of experience in pelvic MRI) was presented with the anatomical images on an Osirix workstation ( 23 ) and asked to prescribe two-dimensional polygonal ROIs on a representative slice for each distinct region larger than 2 cm within all ovarian masses, whether single, bilateral, or multiple, within each subject. Each ROI was then subjectively identified as predominantly cystic or solid based on morphology and intensity in the T 2 -weighted anatomical images. ROIs were copied to the parametric maps, with minor manual adjustments as needed to adjust for motion between the acquisitions and distortion on the DWI images. Histograms were generated for each ROI on all intersecting parametric maps and summary statistics (mean, standard deviation, skewness, and kurtosis) were recorded. Noise-masked pixels were excluded from the analysis. The primary analyses compared the mean value of each parameter in each ROI to histology outcome using generalized estimating equations with exchangeable working correlation structure to account for the potential correlation between multiple ROIs from the same subject. Secondary analyses were performed by considering the solid and cystic ROIs separately. Bonferroni corrections for multiple comparisons were not applied in these exploratory analyses due to the small numbers in the study. For the histogram analysis, we compared the classification accuracy of the ROI mean value alone versus a linear combination of histogram characteristics (mean, standard deviation, skewness, and kurtosis) from each ROI. Linear model coefficients were estimated using logistic regression and the linear combination of the four parameters represents the fitted value for each ROI. The classification accuracy of our model was evaluated using the receiver-operating characteristic (ROC) curve, which was estimated using leave-one-out cross validation (CV) to correct for overfitting. We used the area under the ROC curve (AUC) to compare our fitted model to the classification accuracy for the mean ROI value alone. Confidence intervals for the AUC were estimated by bootstrap. All statistical calculations were performed with the R statistical software environment ( 24 ) and used a 0.05 significance level.

A total of 37 women with adnexal masses gave informed consent and were prospectively and consecutively enrolled in the study. All subjects were scheduled for surgical removal of at least one ovary. One subject was excluded from the analysis because her mass was not ovarian, and two were excluded because they had low malignant potential/borderline ovarian tumors. Thirty-four women were included in the analysis, 12 with malignant ovarian masses and 22 benign. Mean age was 53.6 years (range 34–87). The histologic diagnoses of the 37 women enrolled are detailed in Table 3 . Several of these subjects did not have the full set of parametric data included in the analyses: ten subjects were unable to receive gadolinium-DTPA due to contraindications and therefore had no contrast-related parametric maps (3 malignant, 7 benign); three subjects had diffusion-weighted imaging that was judged unacceptable due to artifacts and therefore had no diffusion-related parametric maps (1 malignant, 2 benign). All 34 subjects had acceptable T 2 -weighted imaging and T 2 maps. The full anatomic and quantitative imaging protocol was completed in an average of 52 minutes (range 42-64). Using the anatomical images, the radiologist prescribed a total of 109 lesion ROIs in the 34 subjects, with either single, bilateral, or multiple ovarian masses. Cystic ROIs were more common: the 22 benign cases included 8 solid ROIs and 41 cystic ROIs, whereas the 12 malignant cases included 20 solid and 40 cystic ROIs. The ROIs were then transferred to all overlapping parametric maps for each of the 15 parameters in Table 2 . Note that the parametric acquisitions had reduced coverage in the slice direction compared to the anatomical images, and therefore not all ROIs were representable in all parametric maps. A total of 1696 combinations of ROI and parametric map were considered for subsequent analyses. Examples of ROIs on anatomical and parametric images are given for a benign case ( Figure 1 ) and a malignant case ( Figure 2 ). The first analysis assessed the mean parametric value within each ROI to identify which MRI parameters were associated with malignancy. Both solid and cystic ROIs were included in this analysis. The results are summarized in Table 4 . Most of the DCE parameters had significantly higher mean values in malignant masses compared to benign masses, including IAUC, IAUC-ref, K trans , k ep , sigmoid EA, and sigmoid EA-ref (p<.001 in all cases). Additionally, the mean values from the T 2 maps were significantly lower in malignant masses compared in benign masses (p=0.028). There were no statistically significant differences with the diffusion-related parameters. To correct for the potentially confounding effect of cystic versus solid ROIs, the analysis was repeated separately for solid ROIs and cystic ROIs. The results from analyzing only the solid ROIs are summarized in Table 5 . Several DCE parameters were significantly different between benign and malignant solid regions, including IAUC-ref, k ep , sigmoid EA-ref, and sigmoid T h (p=0.028, <0.001, 0.025, 0.019, respectively). With the cyst ROIs removed, the T 2 values are higher in malignant solid regions, with significant differences in the T 2 map, the T 2 weighted image signal intensity (T 2 -w) and the normalized signal intensity (T 2 w-ref) (p=0.003, 0.008, <0.001, respectively). No difference was found in ADC or D, but the perfusion fraction F p was significantly lower in malignant solid regions (p=0.043). Even though cystic ROIs generally do not enhance, the predominantly cystic ROIs were analyzed with DCE so that our methods were consistent. Several DCE parameters had higher mean values in malignant cystic regions compared to benign cystic regions, including IAUC (p<0.001), IAUC-ref (p=0.013), K trans (p=0.001), sigmoid EA (p=0.002), and sigmoid EA-ref (p=0.001) (data not shown). No significant differences were found with the T 2 parameters or the diffusion-related parameters. Finally, we compared the classification accuracy of the ROI mean value alone versus a combination of histogram characteristics from each ROI. For this analysis, the data were filtered to select only those ROI/parameter map combinations with reliable distribution metrics: ROI/parameter map combinations that had fewer than 100 pixels after noise thresholding, or less than 33% pixels included after noise thresholding were removed from the analysis. After this filtering, the remaining data contained too few ROIs to adequately assess the DWI and DCE parameters, so the histogram analysis focused solely on the T 2 -related parameters. Figure 3 gives examples of ROIs on T 2 maps and their associated histogram characteristics for benign and malignant cases. Table 6 presents AUC for the mean T 2 imaging parameters and for a linear combination of the mean, standard deviation, skewness and kurtosis for each T 2 imaging parameter. The CV adjusted AUC for a linear combination of the four histogram characteristics for T 2 map increased to 0.90 (95% CI: 0.83, 1.00), compared to an AUC of 0.81 (0.56, 1.00) for the mean alone. This suggests there is additional information for predicting histology outcome beyond the mean of the T 2 map. In contrast, the AUC for T 2 w and T 2 w-ref actually decreased when adding additional histogram characteristics to the model, suggesting that the additional data are simply adding noise to the model and are not associated with benign or malignant masses.

In conclusion, we have demonstrated the feasibility of performing multi-parametric, quantitative MRI for the characterization of ovarian masses, and found several parameters based on DCE-MRI, diffusion imaging, and T 2 relaxometry that were significantly associated with malignancy. The most objective findings were based on the first analysis, which used only ROI means and included all ROIs without subjectively distinguishing between predominantly cystic and solid ROIs. For this analysis, both DCE-MRI and T 2 -based acquisitions produced significant parameters, while diffusion imaging did not. These findings, however, were confounded by composition of the ROIs: predominantly cystic regions generally have longer T 2 values and do not enhance, and purely cystic masses are most always benign. The second analysis, which used only ROI means and only in predominantly solid ROIs, is more relevant to the diagnostic setting where the radiologist is concerned about suspicious solid regions. In this analysis, several parametrics from DCE-MRI, T 2 imaging, and diffusion imaging were associated with malignancy. The DCE-MRI findings were expected and consistent with previous literature ( 16 ), as malignant ovarian masses are known to have increased vascularity. Significant associations were found in parameters derived from each of the three distinct DCE-MRI analysis approaches (area under curve, sigmoid fitting, and pharmacokinetic modeling), suggesting that the diagnostic information is inherent in the data and robust with respect to analysis method. Our finding that quantitative T 2 mapping was associated with ovarian malignancy has not been previously reported. This is, however, consistent with the practice of using hypo- and hyper-intensity on T 2 -weighted imaging as a diagnostic feature ( 14 ), but with the added advantage of being assessed quantitatively. The diffusion parameters performed more poorly than expected, with no association between ADC in either the solid-only or the solid+cyst analysis. This finding was consistent with one prior study ( 18 , 25 ), but inconsistent with several previous reports using 1.5T MRI ( 17 , 26 - 29 ). This may be attributable to our sequence optimization; we scanned with a high resolution (192×192 matrix) while using a large acceleration factor (R=3) to keep the echo train short, and this approach may have led to insufficient signal-to-noise ratio (SNR). Further work is warranted to optimize the diffusion-weighted imaging methods for high-resolution and good SNR at 3T. We did find that the perfusion fraction F p , as estimated by the IVIM technique, was lower in malignant solid regions than in benign solid regions. This finding should be interpreted cautiously, as it is inconsistent with the increased vascularity shown in our DCE-MRI findings. The 3 b-value technique we used to calculate F p was consistent with the originally reported IVIM method ( 19 ), but the F p maps had generally low signal-to-noise, and more b-values should be used in future studies to assess IVIM parameters. The histogram analysis illustrates that other characteristics of the T 2 map distribution beyond the mean provide additional information for discriminating between benign and malignant regions. One explanation may be that malignant masses are typically heterogeneous with a combination of cystic and solid elements, while the majority of benign masses are cystic and homogenous. Even when focusing on only the solid components of both malignant and cystic masses, malignant solid masses have small necrotic regions ( 12 ), whereas the benign solids generally appeared to be more fibrous and uniform. Considering all three analyses, the quantitative T 2 mapping was stronger than T 2 -weighted imaging signal intensity, with or without normalization by a reference ROI. This T 2 mapping has the advantage of removing receive coil sensitivity, T 1 weighting, and proton density effects, and is therefore are more direct reflection of the underlying tissue relaxation rate. Our study has several limitations. First, for the number of parameters investigated and the multiple comparisons performed, the number of subjects enrolled was small. This may decrease the strength of the associations that were seen in this exploratory study, particularly for those in the smallest cohorts. Second, we did not compare the performance of the metrics with conventional interpretation of the anatomic images, as the study was designed to identify potential diagnostic biomarkers rather than assess diagnostic performance. Third, due to variations in spatial coverage between the imaging methods as well as the small study size, we did not have sufficient data to assess the feasibility of using combined parameters (e.g., T 2 and ADC together) in a diagnostic model. Fourth, borderline ovarian tumors comprised 5.4% of patients with ovarian masses who were enrolled in our study (n=2), and they were excluded from the analysis because of the small number. The inclusion of larger numbers of borderline tumors in future trials could be important because there may be distinct differences between the quantitative MRI values in borderline ROIs compared to benign and malignant ROIs. Borderline ovarian tumors are treated surgically, so it would be valuable to differentiate them from benign masses. In conclusion, the development of innovative, quantitative MRI techniques for potential screening and treatment monitoring applications motivates our investigation into quantitative MRI. Utilizing our exploratory data in future studies could contribute to the development of a non-contrast ovarian MRI exam using quantitative T 2 mapping along with conventional image interpretation; this would reduce scan times and cost, and could be useful for potential future studies of new screening modalities. Additionally, based on our exploratory findings, future studies could evaluate the accuracy of a contrast-enhanced MRI study using T 2 mapping methods for the characterization ovarian masses.