Yahui Peng, PhD 2 Yulei Jiang, PhD Tatjana Antic, MD Maryellen L. Giger, PhD Scott E. Eggener, MD Aytekin Oto, MD. Purpose: Materials and Methods:

Transcription

1 Note: This copy is for your personal non-commercial use only. To order presentation-ready copies for distribution to your colleagues or clients, contact us at Validation of Quantitative Analysis of Multiparametric Prostate MR Images for Prostate Cancer Detection and Aggressiveness Assessment: A Cross-Imager Study 1 Original Research n Genitourinary Imaging Yahui Peng, PhD 2 Yulei Jiang, PhD Tatjana Antic, MD Maryellen L. Giger, PhD Scott E. Eggener, MD Aytekin Oto, MD Purpose: Materials and Methods: To validate three previously identified quantitative image features across multiparametric magnetic resonance (MR) images acquired with imagers made by two different manufacturers to differentiate prostate cancer (PC) from normal prostatic tissue and to assess cancer aggressiveness. This study was HIPAA-compliant and approved by the institutional review board. Preoperative 1.5-T multiparametric endorectal MR images of 119 PC patients (dataset A, 71 patients; dataset B, 48 patients) were analyzed, and 265 PC and normal peripheral zone regions of interests (ROIs) were identified through histologic and MR consensus review. The 10th percentile average apparent diffusion coefficient (ADC) value, average ADC value, and skewness of T2-weighted signal-intensity histogram were evaluated with area under the receiver operating characteristic curve (AUC). The image features were combined with a linear discriminant analysis classifier and evaluated both on the image dataset of each type of imager alone (leave-one-patient-out evaluation) and across the datasets (training on one dataset, testing on the other). Spearman correlation coefficient was calculated between the image features and ROI-specific Gleason scores. 1 From the Departments of Radiology (Y.P., Y.J., M.L.G., A.O.), Pathology (T.A.), and Surgery Section of Urology (S.E.E.), University of Chicago, 5841 S Maryland Ave, Chicago, IL Received June 7, 2013; revision requested July 25; revision received October 21; accepted November 8; final version accepted December 9. Supported in part by the U.S. Army Medical Research and Materiel Command Prostate Cancer Research Program through an Idea Development Award (PC093485). Address correspondence to Y.P. ( yhpeng@bjtu.edu.cn). 2 Current address: School of Electronic and Information Engineering, Beijing Jiaotong University, No. 3 Shangyuancun, Haidian District, Beijing, China, (Y.P.). Results: AUC values of the image features combined were (standard error) and on dataset B and dataset A alone, respectively, and and when training on dataset A and testing on dataset B and vice versa, respectively. Spearman correlation coefficients between Gleason scores and the ADC features were between and Conclusion: Consistently across images from datasets A and B, the 10th percentile ADC value, average ADC value, and T2- weighted skewness can distinguish PC from normal-tissue ROIs, and ADC features correlate moderately with ROIspecific Gleason scores. q RSNA, 2014 q RSNA, 2014 Radiology: Volume 271: Number 2 May 2014 n radiology.rsna.org 461

2 Multiparametric magnetic resonance (MR) imaging is currently the best imaging method to identify and characterize prostate cancer (PC), which is the second leading cause of cancer-related death in men in the United States (1,2). The current practice of PC diagnosis based on transrectal ultrasonography-guided core needle biopsy has limitations that include the invasiveness of biopsy, likelihood of missing anterior tumors, and inaccuracy in assessment of the aggressiveness of PC or Gleason score (3). Multiparametric MR imaging provides anatomic information in T2-weighted images and functional information in images that are diffusion weighted (DW), dynamic contrast material enhanced (DCE), and spectroscopic and has the potential to help improve the detection of PC and the differentiation of clinically significant PC from indolent tumors (2,4,5). Interpretation of multiparametric MR prostate images is a challenge for radiologists because of an overwhelming Advances in Knowledge nn The 10th percentile and the average apparent diffusion coefficient (ADC) values and T2-weighted signal-intensity skewness are consistently effective (area under the receiver operating characteristic curve [AUC], ) for differentiating prostate cancer (PC) foci from normal peripheral zone regions of interests (ROIs) in multiparametric endorectal MR images acquired from 1.5-T imagers made by two different manufacturers. nn The 10th percentile ADC value is consistently more effective than the average ADC value to differentiate PC foci from normal peripheral-zone tissue (AUCs, 0.92 vs 0.89 and 0.89 vs 0.87 on two independent image datasets). nn The 10th percentile and the average ADC values consistently correlate moderately with lesionspecific Gleason scores (r between and 20.34). amount of data that need to be analyzed and the lack of a standardized interpretation approach that would lead to reproducible results. Because T2-weighted and DW MR images are the most important sequences that are helpful in PC detection, efforts have been made for quantitative image analysis of these images by using computer-aided diagnosis (CAD) (6 9). On the basis of data from MR imagers from a single manufacturer, we previously reported (9) that the 10th percentile apparent diffusion coefficient (ADC) value within a tumor region of interest (ROI), the average ADC value, and T2-weighted signal-intensity skewness within an ROI were effective image features for differentiation of PC foci from normal prostatic tissue, and that both of the ADC features correlated moderately with the Gleason score of a tumor. However, one of the challenges for widespread use of MR imaging for PC is large variations in images acquired with imagers made by different manufacturers. The purpose of this study was to validate the three previously identified quantitative image features across multiparametric MR images acquired with imagers made by two different manufacturers (Signa, GE Healthcare, Waukesha, Wis; Achieva, Philips Healthcare, Eindhoven, the Netherlands) in order to differentiate PC from normal prostatic tissue and to assess cancer aggressiveness. Materials and Methods Patients This study was retrospective, compliant with the Health Insurance Portability and Implication for Patient Care nn The quantitative ADC and T2-weighted image features evaluated in this study can be used consistently in multiparametric endorectal MR images acquired with 1.5-T imagers made by two different manufacturers for the differentiation of PC from normal-tissue ROIs and for assessment of correlation strength with lesion-specific Gleason scores. Accountability Act, and approved by our institutional review board with a waiver of written informed patient consent. We searched the radiologic image archive at our institution and identified 139 consecutive patients between July 2007 and May 2010 by using the following inclusion criteria: (a) patient had biopsy-proved prostate adenocarcinoma, (b) patient underwent multiparametric MR examination with an endorectal coil (including T1-weighted, T2-weighted, DW MR imaging, and DCE MR sequences), and (c) patient underwent radical prostatectomy. Twenty patients were excluded because of the following: patient received radiation therapy before the MR examination (n = 1), the radical prostatectomy was performed before the MR examination (n = 2), prostatectomy histologic slides were not available for review (n = 6), the MR examination was performed with 3-T imagers (n = 5), DW MR image data were missing (n = 1), or DCE MR image data were missing (n = 5). A total of 119 patients were included in this study. Among them, 59.7% (71 of 119; dataset A) were imaged with 1.5-T imagers (Signa; GE Healthcare) between July 2007 and October 2008, and 40.3% (48 of 119; dataset B) were imaged with 1.5-T imagers (Achieva; Published online before print /radiol Content code: Radiology 2014; 271: Abbreviations: ADC = apparent diffusion coefficient AUC = area under the ROC curve CAD = computer-aided diagnosis DCE = dynamic contrast enhanced DW = diffusion weighted PC = prostate cancer ROC = receiver operating characteristic ROI = region of interest Author contributions: Guarantors of integrity of entire study, Y.P., Y.J., A.O.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; literature research, Y.P., Y.J., M.L.G., A.O.; clinical studies, A.O.; experimental studies, Y.P., M.L.G., A.O.; statistical analysis, Y.P., Y.J., M.L.G., A.O.; and manuscript editing, all authors Conflicts of interest are listed at the end of this article. 462 radiology.rsna.org n Radiology: Volume 271: Number 2 May 2014

3 Philips Healthcare) between March 2008 and March Dataset B was reported in a previous study (9). Patient age and prostate-specific antigen at diagnosis are summarized in Table 1. MR Image Acquisition An endorectal coil (Medrad; Bayer Healthcare, Warrendale, Pa) inflated with air and a phased-array surface coil were used for all MR images except for DCE MR sequences in dataset A, which were acquired with only a phased-array surface coil. Immediately before the MR examination, 1 mg glucagon (Glucagon; Lilly, Indianapolis, Ind) was injected intramuscularly to reduce peristalsis of the rectum. Axial, coronal, and sagittal T2-weighted images, axial T1-weighted images, axial free-breathing DW MR images, and axial free-breathing DCE MR images of the entire prostate were acquired. Orientation of axial images was perpendicular to the rectal wall and guided by sagittal images. Acquisition of DCE MR images started 30 seconds before intravenous administration of 0.1 mmol/kg gadodiamide (Omniscan; GE Healthcare) followed by a 20-mL saline flush at a rate of 2.0 ml/sec. Additional image acquisition parameters are detailed in Tables 2 and 3 for datasets A and B, respectively. Histologic Analysis and MR Correlation For this study, archived tissue sections from the entire prostate were re-evaluated for use in the histologic analysis and MR correlation. All tissue sections of prostatectomy specimens from the 119 patients were retrieved from the Department of Pathology tissue archive of histologic and MR correlation analysis. Prostatectomy specimens were cut serially into 4-mm-thick blocks from the apex to the base in transverse planes and then halved or quartered depending on size. After fixation in 5% buffered formalin, each block was processed and embedded into paraffin. Finally, 4 5 mm microtomed sections were obtained and stained with hematoxylin-eosin. A genitourinary pathologist (T.A., 8 years of experience in genitourinary pathologic analysis) and a radiologist Table 1 Patient and ROI Characteristics Parameter Dataset A Dataset B No. of patients Median patient age (y)* 61.0 (47 75) 62.5 (44 73) Median PSA (ng/ml)* 5.3 ( ) 7.0 ( ) Total no. of ROIs No. of normal ROIs 59/161 (36.6) 43/104 (41.3) No. of cancer ROIs 102/161(63.4) 61/104 (58.7) No. of peripheral zone cancer ROIs 94/102 (92.2) 47/61 (77.0) No. of central gland cancer ROIs 8/102 (7.8) 14/61 (23.0) Median ROI size (mm 2 )* 55.0 ( ) 50.4 ( ) Patient Gleason score 6 23/102 (22.5) 14/61 (23.0) 7 57/102 (55.9) 32/61 (52.5) 8 10/102 (9.8) 9/61 (14.8) 9 12/102 (11.8) 6/61 (9.8) Note. Unless otherwise indicated, data in parentheses are percentages. PSA = prostate-specific antigen. * Data in parentheses are the range. PSA data were missing in one patient. PSA data were missing in six patients. (A.O., 9 years of experience in prostate MR imaging) established the reference standard for PC and normaltissue foci on MR images through a systematic histologic and MR correlative review. The pathologist identified all distinct tumor foci larger than approximately 5 mm in diameter after a review of all prostatectomy tissue sections. Then, by consensus, the radiologist manually outlined ROIs of the corresponding tumor foci on MR images that best correlated with the histologic findings. For tumor foci that were not discernible on MR images, their locations on MR images were determined on the basis of the spatial relationship to other identifiable anatomic landmarks (eg, urethra, ejaculatory ducts, and benign prostatic hyperplasia). The radiologist drew ROIs on MR images that best aligned with the tumor that was identified on the specimen by the pathologist. In each case, tumor ROIs of the peripheral zone and/or central gland (which consists of the central and transition zones) were outlined if they were present. A normal-tissue focus was also outlined in the peripheral zone in locations that the pathologist indicated as normal (the pathologist considered high-grade prostatic intraepithelial neoplasia, benign prostatic hyperplasia, and other benign abnormalities abnormal), unless normal region could not be found on histologic slides (22 patients). A total of 265 ROIs were outlined, among which 102 (38.5%) were from normal peripheral zone tissue and 163 (61.5%) were from PC tumors (Table 1). Although almost all ROIs were outlined in a T2- weighted image section, seven cancer ROIs in six cases were outlined in a DW MR image section and six other cancer ROIs in six cases were outlined in a DCE MR image section because those ROIs correlated better with histologic findings in those images than in the T2-weighted images. The pathologist also assigned a Gleason score specifically to each cancer ROI (Table 1). Transfer of ROIs between MR Images The manually drawn ROIs were transferred to other MR image sequences to facilitate the analysis in those images. First, assuming no patient motion throughout the entire multiparametric MR examination, the ROIs were automatically transferred from the sequence where the ROIs were drawn manually Radiology: Volume 271: Number 2 May 2014 n radiology.rsna.org 463

4 Table 2 MR Image Acquisition Parameters for Dataset A No. of Signals Acquired Flip Angle (degrees) Sequence Sequence Type TR (msec) TE (msec) Field of View (mm)* Matrix In-plane Resolution (mm 2 ) Section Thickness (mm) T2 weighted (axial) FSE , T2 weighted (sagittal) FSE , T2 weighted (coronal) FSE , DW (axial) EPI SE , 4 DCE (axial) FFE Note. Dataset A was GE Healthcare. An array spatial sensitivity encoding technique (parallel imaging) factor of two was used in all sequences. Total imaging acquisition time was approximately 45 minutes. EPI SE = echo-planar imaging spin echo, FFE = fast field echo, FSE = fast spin echo, TE = echo time, TR = repetition time. * Listed is the length of one side of a square field of view. Diffusion weighting factor, b values, were 0 and 500 sec/mm 2 in one patient (1.4%; one of 71) and 0 and 1000 sec/mm 2 in 70 patients.. Four cases were excluded because their DCE MR images were missing. Approximately DCE MR images were acquired in 5 7 minutes at a temporal resolution of 5 12 seconds. to other sequences. Then, the study radiologist reviewed the transferred ROIs in all cases to confirm the accuracy of the ROI locations. In cases where misalignments were visually obvious, the ROI locations were manually shifted without modifying their size or shape. Image Feature Analysis Three image features were previously identified as potentially effective in the task of differentiating PC foci from normal peripheral zone ROIs: 10th percentile ADC value, average ADC value, and the skewness of T2-weighted signalintensity histogram (9). Other image features that were previously reported in the literature, including DCE image features, were not among the strongest identified in the previous study (9), which is generally in agreement with current clinical experience and literature (10,11). These three image features are briefly summarized here. Pixel-wise ADC values were calculated from DW MR images by using a monoexponential model, ADC 1nDW 1nDW, b 0 b = where DW 0 and DW b are DW MR signal intensities with DW of 0 and b values, respectively, and 1n represents natural logarithm (12,13). When more than two b values were available, we estimated the pixel-wise ADC value by using a linear least-squares fit to DW MR image data of all available b values (on rare occasions it was necessary to set the pixel-wise ADC value to 0 when a fit yielded a negative value). The average and 10th percentile ADC values were calculated from the histogram of pixel-wise ADC values within each ROI. The average ADC value was expected to help characterize restricted water diffusion in tumors, and the 10th percentile ADC value was expected to help characterize sparse tumors (tumor intermixed with normal tissue) more effectively than the average ADC value. The skewness of T2-weighted signal intensity was calculated from the 464 radiology.rsna.org n Radiology: Volume 271: Number 2 May 2014

5 Table 3 MR Image Acquisition Parameters for Dataset B Flip Angle (degrees) No. of Signals Acquired Sequence Sequence Type TR (msec) TE (msec) Field of View (mm)* Matrix In-plane Resolution (mm 2 ) Section Thickness (mm) T2-weighted (axial) FSE , T2-weighted (sagittal) FSE , , T2-weighted (coronal) FSE , , DW (axial) FSE EPI DCE (axial) FFE Note. Dataset B was Phillips Healthcare. An effective sensitivity encoding (parallel imaging) factor of two was used in all sequences. Total imaging acquisition time was approximately 45 minutes. FFE = fast field echo, FSE = fast spin echo, FSE EPI = FSE with echo-planar imaging readout, TE = echo time, TR = repetition time. * Listed is the length of one side of a square field of view. Diffusion weighting factor, b values, were 0, 50, 200, 1500, and 2000 sec/mm 2 in 24 patients (50%; 24 of 48), and 0 and 1000 sec/mm 2 in the other 24 patients. Approximately DCE MR images were acquired in 5 10 minutes at a temporal resolution of 3 6 seconds. histogram of T2-weighted signal intensity within each ROI. It was defined as where S denotes the summation operator, S i denotes T2-weighted signal intensity of the i pixel, m and s denote the mean and standard deviation of the T2-weighted signal-intensity histogram, and N denotes the total number of pixels within the ROI. Positive skewness would indicate that the ROI contains more dark pixels than bright pixels, which is typical in tumors, and negative or zero skewness would indicate fewer or equal number of dark pixels compared with bright pixels, which is typical in normal-tissue ROIs. MR images were retrieved from a picture archiving and communication system, and all image analyses were performed offline by using in-house computer software written in the Python programming language (version 2.6.5; Statistical Analysis Receiver operating characteristic (ROC) analysis was used to characterize the effectiveness of the image features to differentiate PC foci from normal-tissue ROIs (14). Maximum-likelihood estimated proper binormal ROC curves were obtained (15), and the area under the ROC curve (AUC) was used as a figure of merit (16). Both per-roi and per-patient analyses were performed. In the per-roi analysis, each ROI was used as an independent unit of analysis. In the per-patient analysis, two or more cancer (and, independently, normal-tissue) ROIs within a single patient were combined. Minimum ADC features and maximum T2-weighted skewness of the ROIs (all indicative of PC) were used as the combined features. We combined the three image features by using linear discriminant analysis classifiers (17). For analysis of dataset A or dataset B alone, a leave-one-patient-out cross-validation method was used to separate training from testing of the linear discriminant 3, Radiology: Volume 271: Number 2 May 2014 n radiology.rsna.org 465

6 analysis classifier; ROIs from each patient were used, in turn, to test the linear discriminant analysis classifier while ROIs from all other patients were used to train the classifier, and subsequently, ROC analysis was conducted on the test results of all patients in aggregate (18). Furthermore, crossdataset validation was conducted by training a linear discriminant analysis classifier on dataset A and testing it on dataset B and vice versa. ROC curves were compared statistically between the two image datasets in terms of AUC (16). The Spearman rank-order correlation coefficient was calculated to characterize the correlation strength between each image feature and the ROI-specific Gleason score. The Pearson correlation coefficient was used to characterize the correlation strength between two image features (19). All statistical tests were two sided, and P value less than.05 was the critical value that indicated statistical significance. We calculated 10 P values for comparisons of AUC values and 14 P values for correlation coefficients; application of Bonferroni correction for multiple comparisons would cause the critical value for statistical significance to be adjusted to P value less than.005 for the AUC comparisons and P less than.004 for the correlation coefficients, or P value less than.002 for all. Statistical analyses were performed by using inhouse computer software (Python), and ROC analysis was performed by using software developed by Metz (20). Results Two example cases of T2-weighted MR images and ADC maps obtained by using MR imagers from the two manufacturers with the corresponding histologic tissue sections are shown in Figures 1 and 2, respectively. In both of these cases, ROIs were outlined on the T2-weighted MR images and then automatically transferred to the ADC maps. The AUC values of the 10th percentile ADC value, average ADC value, and T2-weighted skewness in differentiating PC ROIs from normal peripheral zone Figure 1 Figure 1: A, Example T2-weighted MR image, and, B, ADC map in a 69-year-old patient in dataset A. The patient s tumor had a Gleason score of = 7, outlined in A and then automatically transferred to the ADC map (B). C, D, Corresponding dissected tissue sections show the same tumor (arrow). (Total magnification 32.5 in C and 350 in D.) MR images are enlarged to 300% of the original size, and window and level are adjusted individually for better visualization. tissue, separately on datasets A and B, are shown in Table 4. The 10th percentile ADC value was the most effective on both datasets, with AUC values of (standard error) for dataset A and for dataset B (per-roi analysis) and and (per-patient analysis), respectively. The 10th percentile and the average ADC values were highly correlated (Fig 3). Both the 10th percentile and the average ADC values were slightly greater for dataset A than for dataset B, but the overlap between PC and normal-tissue ROIs appeared to be similar between the two image datasets (Fig 3). The 10th percentile ADC value and T2-weighted skewness were not correlated (Fig 4), which indicates that T2-weighted skewness provided additional independent information. The distributions of these feature values in Figure 4 were qualitatively similar between the two image datasets. Combining the three image features with a linear discriminant analysis classifier separately for both datasets by using the leave-one-patient-out cross validation method yielded AUC values of and for datasets A and B, respectively (Table 4). When the linear discriminant analysis classifier was trained on dataset B and tested on dataset A, AUC values of (per ROI) and (per patient) were obtained. When the classifier was trained on dataset A and tested on dataset B, AUC values of (per ROI) and (per patient) were obtained (Fig 5). AUC value difference 466 radiology.rsna.org n Radiology: Volume 271: Number 2 May 2014

7 Figure 2 ROI-specific Gleason scores (Fig 6). The correlation appeared to be similar between the two image datasets, although some differences were apparent for PC ROIs with Gleason scores of 8 or 9 (Fig 6). T2-weighted skewness did not appear to correlate with ROI-specific Gleason scores, which yielded Spearman rank-order correlation coefficients of (P =.65) and (P =.59) for datasets A and B, respectively. Figure 2: A, Example T2-weighted MR image and, B, ADC map in a 65-year-old patient in dataset B. The patient had a Gleason score 4+3 = 7 tumor, outlined in A and then automatically transferred to the ADC map (B). C, D, Corresponding dissected tissue sections show the same tumor (arrow). (Total magnification, 32.5 in C and 350 in D.) MR images are enlarged to 300% of the original size and with window and level adjusted individually for better visualization. Table 4 AUC Values of the Image Features for Distinguishing PC Foci from Normal Peripheral Zone Tissue ROIs Image Feature Per-ROI Analysis Per-Patient Analysis Dataset A Dataset B P Value* Dataset A Dataset B P value* 10th percentile ADC Average ADC T2-weighted skewness All features combined Note. Data shown are maximum-likelihood estimate AUC 6 standard error. * P value is in terms of the difference in AUC values between the two datasets. The three features are combined with a linear discriminant analysis classifier and evaluated with the leave-one-patient-out cross-validation method. was not statistically significant in both per-roi and per-patient analyses (P =.05 for both). Both the 10th percentile and the average ADC values were moderately and negatively correlated with the Discussion Validation of quantitative image features across MR imagers made by different manufacturers is important because proprietary pulse sequences, signal processing, and image processing can affect the appearance of images and, consequently, affect quantitative image features and their effectiveness in CAD image analysis. In this study, we evaluated the effectiveness of three image features in the differentiation of PC and normal-tissue ROIs, and in their correlation with lesion-specific Gleason scores on images acquired from MR imagers made by two different manufacturers. Our results indicate that the 10th percentile ADC value, the average ADC value, and T2-weighted signalintensity skewness are consistently robust image features across both the 1.5-T image datasets and can be used for prostate MR CAD development. We estimated pixel-wise ADC values with a monoexponential model. Given the different DW MR acquisition protocols between the two image datasets (different b values and different numbers of b values), it is interesting to note that the discriminant ability of the ADC features was not significantly affected. This is consistent with previous reports that indicate that although the use of different b values can cause systematic changes in ADC values, the diagnostic effectiveness of the ADC value is not necessarily degraded (21,22). The 10th percentile ADC value was consistently more effective than the average ADC value for differentiation of PC ROI from normal peripheral zone tissue. This is probably because the Radiology: Volume 271: Number 2 May 2014 n radiology.rsna.org 467

8 Figure 3 Figure 3: Scatterplots of the 10th percentile and the average ADC values for normal (green circles) and prostate cancer (red squares) ROIs for, A, image dataset A and, B, image dataset B. The Pearson correlation coefficient and the number of ROIs are also shown. In four cancer ROIs in dataset A, the 10th percentile ADC value was set to 0 because more than 10% of the pixels in the ROIs had pixel-wise ADC value set to 0. Figure 4 Figure 4: Scatterplots of the 10th percentile ADC value and T2-weighted signal-intensity histogram skewness for normal (green circles) and prostate cancer (red squares) ROIs for, A, image dataset A and, B, image dataset B. The Pearson correlation coefficient and the number of ROIs are also shown. In four cancer ROIs in dataset A, the 10th percentile ADC value was set to 0 because more than 10% of the pixels in the ROIs had pixel-wise ADC value set to 0. 10th percentile ADC value is more representative of the most aggressive and densely packed portions of the lesions, between which normal tissue may also exist (12). Further studies are needed to confirm this. The effectiveness of T2-weighted signal-intensity skewness, characterized by the AUC value, is apparently not as consistent between the two image datasets as that of the two ADC image features. There are two possible reasons 468 radiology.rsna.org n Radiology: Volume 271: Number 2 May 2014

9 Figure 5 for this. First, the T2-weighted images in dataset B were acquired with a larger number of averages and had better signal-to-noise ratio than those in dataset A. Second, the T2-weighted images in dataset A often had stronger inhomogeneity in sensitivity across the field of view than those in dataset B. These differences could have caused degradation in the effectiveness of T2-weighted signal-intensity skewness on dataset A. It is encouraging that the three image features, combined with a linear discriminant analysis classifier, appear to be consistently effective between the two datasets for distinguishing between PC and normal-tissue ROIs. The crossvalidation AUC values of and , which were obtained when the linear discriminant analysis classifier was trained on one image dataset and tested on the other, are consistent with the leave-one-out AUC values obtained on each image dataset alone (Table 4). Given obvious differences in DW MR image acquisition and DW MR and T2-weighted image appearance between the two datasets, the results suggest that these image features are robust across the MR imagers. The statistically significant, moderate, and negative correlation between lesion-specific Gleason score and the Figure 5: Estimated proper binormal (solid curves) and empirical (dotted curves), crossvalidation ROC curves in the per-roi analysis. Red curves: training the linear discriminant analysis classifier on dataset B and testing it on dataset A; green curves: training the linear discriminant analysis on dataset A and testing it on dataset B. The values of AUC are also shown and their difference is not statistically significant (P =.05). The results of the per-patient analysis are similar (not shown). 10th percentile and the average ADC values are consistent between the two image datasets. Lower ADC values are consistently associated with higher lesion-specific Gleason scores. The apparent inconsistencies for high-grade (Gleason score 8 or 9) tumors are probably the result of small number of patients with high-grade tumors. This observation has been previously noted in the literature (23 26). It has been hypothesized (27) that tissue that is dense with cells limits intercellular space, which restricts water molecule diffusion and, thus, results in reduced ADC values. For PC, high-grade tumors are associated with poorly differentiated and often packed epithelial cells compared with low-grade tumors that have at least some individual glandular structures, which preserve some (albeit reduced) intercellular space (28). This may be a reason for the ADC value to be negatively correlated with the Gleason score. However, large interpatient variations are expected to cause overlap in the ADC values between high-grade and low-grade tumors and may explain the moderate correlation strength (23 27). This study has several limitations. First, it is a retrospective study from a single institution with a limited number of patients. Second, patient selection bias exists because all patients had PC and had undergone prostatectomy. Further, normal-tissue ROIs were identified from patients who had PC. This clustering of cancer and normal-tissue ROIs within a single patient could have affected our statistical analysis, which therefore requires further validation. However, these biases are common limitations to many contemporary studies of prostate MR imaging because prostate MR imaging is clinically performed only on select patients. Third, this study focused on differentiation of PC from normal tissue in the peripheral zone. Differentiation of PC from normal tissue in the central gland and from benign prostatic hyperplasia and other benign abnormalities will need to be investigated in the future. Fourth, histologic analysis was done on conventional (ie, dissected) rather than whole-mount sections. Future studies on whole-mount sections may improve the accuracy of the MR histologic correlation analysis of lesions. Fifth, the ADC value was estimated from various DW MR image acquisition protocols in the absence of a consensus in the literature on how to optimize DW MR image acquisition. This variation in the acquisition protocol is expected to have affected estimated ADC values, but it does not appear to have degraded the diagnostic effectiveness of ADC values. Finally, there are uncertainties in the selection of ROIs, and small cancer foci (,5 mm) were not included in this study. Further studies will be needed to address these limitations. In conclusion, this validation study of quantitative multiparametric MR image features shows that 10th percentile ADC value, average ADC value, and T2-weighted signal-intensity skewness are consistently effective across MR imagers made by two manufacturers to distinguish PC from normal-tissue ROIs, and that the 10th percentile and the average ADC values are significantly, moderately, and negatively correlated with lesion-specific Gleason scores, which is consistent across imagers made by two manufacturers. Quantitative image analysis has the Radiology: Volume 271: Number 2 May 2014 n radiology.rsna.org 469

10 Figure 6 Figure 6: Boxplots that show correlation between lesion specific Gleason scores and the 10th percentile ADC value (top), between lesion specific Gleason scores and the average ADC value (bottom), for dataset A (left), and for dataset B (right). The red horizontal lines denote medians, the boxes denote interquartile (the 25th percentile to the 75th percentile) range, and the data points outside the whiskers denote outliers. The Spearman correlation coefficient and the number of ROIs, n, are also shown. Normal ROIs are plotted for comparison only (not included in calculation of correlation coefficients). (In four cancer ROIs in dataset A, the 10th percentile ADC value was set to zero because more than 10% of the pixels in the ROIs had pixel-wise ADC value set to zero.) potential to assist radiologists to better interpret multiparametric prostate MR images. Acknowledgement: This work was supported in part by the U.S. Army Medical Research and Material Command Prostate Cancer Research Program through an Idea Development Award (PC093485). Disclosures of Conflicts of Interest: Y.P. No relevant conflicts of interest to disclose. Y.J. No relevant conflicts of interest to disclose. T.A. No relevant conflicts of interest to disclose. M.L.G. Financial activities related to the present article: none to disclose. Financial activities not related to the present article: travel reimbursement and board membership from the International Society for Optics and Photonics board; board membership on Board of Quantitative Insights; receives money for patents through university s Tech office; author and author s institution receives royalties from IP licensed through university s Tech office; author and author s institution receives money from stock and stock options through the university s Tech office. Other relationships: none to disclose. S.E.E. No relevant conflicts of interest to disclose. A.O. Financial activities related to the present article: none to disclose. Financial activities not related to the present article: grant from Philips Healthcare; payment for development of educational presentations from Philips Healthcare; participation in one GE Healthcare advisory board meeting on prostate cancer. Other relationships: none to disclose. References 1. Siegel R, Naishadham D, Jemal A. Cancer statistics, CA Cancer J Clin 2012;62(1): Hoeks CM, Barentsz JO, Hambrock T, et al. Prostate cancer: multiparametric MR imaging for detection, localization, and staging. Radiology 2011;261(1): Moyer VA; U.S. Preventive Services Task Force. Screening for prostate cancer: U.S. Preventive Services Task Force recommendation statement. Ann Intern Med 2012;157(2): Kozlowski P, Chang SD, Jones EC, Berean KW, Chen H, Goldenberg SL. Combined diffusion-weighted and dynamic contrastenhanced MRI for prostate cancer diagnosis correlation with biopsy and histopathology. J Magn Reson Imaging 2006;24(1): Tanimoto A, Nakashima J, Kohno H, Shinmoto H, Kuribayashi S. Prostate cancer screening: the clinical value of diffusionweighted imaging and dynamic MR imaging in combination with T2-weighted imaging. J Magn Reson Imaging 2007;25(1): Vos PC, Barentsz JO, Karssemeijer N, Huisman HJ. Automatic computer-aided detection of prostate cancer based on multiparametric magnetic resonance image analysis. Phys Med Biol 2012;57(6): Puech P, Betrouni N, Makni N, Dewalle AS, Villers A, Lemaitre L. Computer-assisted diagnosis of prostate cancer using DCE-MRI data: design, implementation and preliminary results. Int J CARS 2009;4(1): radiology.rsna.org n Radiology: Volume 271: Number 2 May 2014

11 8. Niaf E, Rouvière O, Mège-Lechevallier F, Bratan F, Lartizien C. Computer-aided diagnosis of prostate cancer in the peripheral zone using multiparametric MRI. Phys Med Biol 2012;57(12): Peng Y, Jiang Y, Yang C, et al. Quantitative analysis of multiparametric prostate MR images: differentiation between prostate cancer and normal tissue and correlation with Gleason score a computer-aided diagnosis development study. Radiology 2013;267(3): Tretiakova M, Antic T, Binder D, et al. Microvessel density is not increased in prostate cancer: digital imaging of routine sections and tissue microarrays. Hum Pathol 2013;44(4): Oto A, Kayhan A, Jiang Y, et al. Prostate cancer: differentiation of central gland cancer from benign prostatic hyperplasia by using diffusion-weighted and dynamic contrast-enhanced MR imaging. Radiology 2010;257(3): Langer DL, van der Kwast TH, Evans AJ, et al. Intermixed normal tissue within prostate cancer: effect on MR imaging measurements of apparent diffusion coefficient and T2 sparse versus dense cancers. Radiology 2008;249(3): Chenevert TL, Galbán CJ, Ivancevic MK, et al. Diffusion coefficient measurement using a temperature-controlled fluid for quality control in multicenter studies. J Magn Reson Imaging 2011;34(4): Metz CE. ROC methodology in radiologic imaging. Invest Radiol 1986;21(9): Metz CE, Pan X. Proper binormal ROC curves: Theory and maximum-likelihood estimation. J Math Psychol 1999;43(1): Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143(1): Fukunaga K. Introduction to statistical pattern recognition. 2nd ed. Boston, Mass: Academic Press, 1990; Fukunaga K. Introduction to statistical pattern recognition. 2nd ed. Boston, Mass: Academic Press, 1990; Riffenburgh RH. Statistics in medicine. 2nd ed. Amsterdam, the Netherlands: Elsevier Academic Press, 2006; Metz CE. ROC software. uchicago.edu. Accessed June 25, Thörmer G, Otto J, Reiss-Zimmermann M, et al. Diagnostic value of ADC in patients with prostate cancer: influence of the choice of b values. Eur Radiol 2012;22(8): Peng Y, Jiang Y, Antic T, et al. Apparent diffusion coefficient (ADC) for prostate cancer imaging: the impact of the b-values. AJR Am J Roentgenol (in press). 23. Tamada T, Sone T, Jo Y, et al. Apparent diffusion coefficient values in peripheral and transition zones of the prostate: comparison between normal and malignant prostatic tissues and correlation with histologic grade. J Magn Reson Imaging 2008;28(3): Woodfield CA, Tung GA, Grand DJ, Pezzullo JA, Machan JT, Renzulli JF 2nd. Diffusionweighted MRI of peripheral zone prostate cancer: comparison of tumor apparent diffusion coefficient with Gleason score and percentage of tumor on core biopsy. AJR Am J Roentgenol 2010;194(4):W316 W Turkbey B, Shah VP, Pang Y, et al. Is apparent diffusion coefficient associated with clinical risk scores for prostate cancers that are visible on 3-T MR images? Radiology 2011;258(2): Verma S, Rajesh A, Morales H, et al. Assessment of aggressiveness of prostate cancer: correlation of apparent diffusion coefficient with histologic grade after radical prostatectomy. AJR Am J Roentgenol 2011;196(2): Itou Y, Nakanishi K, Narumi Y, Nishizawa Y, Tsukuma H. Clinical utility of apparent diffusion coefficient (ADC) values in patients with prostate cancer: can ADC values contribute to assess the aggressiveness of prostate cancer? J Magn Reson Imaging 2011;33(1): Epstein JI, Netto GJ. Biopsy interpretation of the prostate. 4th ed. Philadelphia, Pa: Lippincott Williams & Wilkins, 2007; Radiology: Volume 271: Number 2 May 2014 n radiology.rsna.org 471