M e d i c a l P hy s i c s a n d I n f o r m a t i c s • O r i g i n a l R e s e a r c h Turkbey et al. Automated Prostate Segmentation on MRI
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Medical Physics and Informatics Original Research
Fully Automated Prostate Segmentation on MRI: Comparison With Manual Segmentation Methods and Specimen Volumes Baris Turkbey 1 Sergei V. Fotin2 Robert J. Huang1 Yin Yin2 Dagane Daar 1 Omer Aras1 Marcelino Bernardo1 Brian E. Garvey 1 Juanita Weaver 1 Hrishikesh Haldankar 2 Naira Muradyan2 Maria J. Merino 3 Peter A. Pinto 4 Senthil Periaswamy 2 Peter L. Choyke1 Turkbey B, Fotin SV, Huang RJ, et al. Keywords: fully automated segmentation, MRI, prostate, segmentation, volume DOI:10.2214/AJR.12.9712 Received July 21, 2012; accepted after revision September 26, 2012. S. Periaswamy is a shareholder in iCAD. 1
Molecular Imaging Program, National Cancer Institute, National Institutes of Health, 10 Center Dr, MSC 1182, Bldg 10, Rm B69, Bethesda, MD 20892-1088. Address correspondence to P. L. Choyke ([email protected]).
2
iCAD Inc., Nashua, NH.
3
Laboratory of Pathology, NCI, NIH, Bethesda, MD.
4
Urologic Oncology Branch, NCI, NIH, Bethesda, MD.
Supplemental Data Available online at www.ajronline.org. WEB This is a web exclusive article. AJR 2013; 201:W720–W729 0361–803X/13/2015–W720 © American Roentgen Ray Society
W720
OBJECTIVE. The objective of our study was to compare calculated prostate volumes derived from tridimensional MR measurements (ellipsoid formula), manual segmentation, and a fully automated segmentation system as validated by actual prostatectomy specimens. MATERIALS AND METHODS. Ninety-eight consecutive patients (median age, 60.6 years; median prostate-specific antigen [PSA] value, 6.85 ng/mL) underwent triplane T2weighted MRI on a 3-T magnet with an endorectal coil while undergoing diagnostic workup for prostate cancer. Prostate volume estimates were determined using the formula for ellipsoid volume based on tridimensional measurements, manual segmentation of triplane MRI, and automated segmentation based on normalized gradient fields cross-correlation and graph-search refinement. Estimates of prostate volume based on ellipsoid volume, manual segmentation, and automated segmentation were compared with prostatectomy specimen volumes. Prostate volume estimates were compared using the Pearson correlation coefficient and linear regression analysis. The Dice similarity coefficient was used to quantify spatial agreement between manual segmentation and automated segmentation. RESULTS. The Pearson correlation coefficient revealed strong positive correlation between prostatectomy specimen volume and prostate volume estimates derived from manual segmentation (R = 0.89–0.91, p < 0.0001) and automated segmentation (R = 0.88–0.91, p < 0.0001). No difference was observed between manual segmentation and automated segmentation. Mean partial and full Dice similarity coefficients of 0.92 and 0.89, respectively, were achieved for axial automated segmentation. CONCLUSION. Prostate volume estimates obtained with a fully automated 3D segmentation tool based on normalized gradient fields cross-correlation and graph-search refinement can yield highly accurate prostate volume estimates in a clinically relevant time of 10 seconds. This tool will assist in developing a broad range of applications including routine prostate volume estimations, image registration, biopsy guidance, and decision support systems.
S
egmentation of the prostate from surrounding tissue on MRI is useful for a variety of clinical purposes including determination of prostate volume, prostate-specific antigen (PSA) density, registration of MRI with other modalities such as ultrasound and PET, and imagingguided biopsy and therapy. Prostate volumes are correlated with lower urinary tract symptoms and are also relevant to decisions regarding the feasibility of brachytherapy and surgery [1–10]. The current methods of estimating prostate volume include physician estimation based on a digital rectal examination (DRE) and imaging measurements based on ultrasound, CT, or MRI. The imaging measurements generally rely on triplanar linear measurements to calculate prostate volume
using the following formula for the volume of an ellipsoid: (length × width × height × 0.52), which assumes that the prostate has a regular ovoid shape [11]. The DRE is often inaccurate because it is based on a subjective estimate of the examiner. Prostate volume determinations based on the ellipsoid formula are often inaccurate because the shape of the prostate varies dramatically [12]. As benign prostatic hyperplasia develops, the prostate evolves from a coneshaped organ to a more spheric organ that often includes an eccentrically enlarged median lobe that is not accounted for by the ellipsoid formula. Manual planimetry based on imaging sections is more accurate but is time-con-
AJR:201, November 2013
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI suming and requires expertise [13–17]. Automating this process could result in highly accurate yet timely prostate volume determinations that could be incorporated routinely into clinical interpretations. In addition to rapid prostate volume determinations, there are other reasons to justify automated segmentation systems. For multiparametric MRI, it is important to align all images (T2, diffusion-weighted, dynamic contrast-enhanced images) to offset misregistration due to patient motion. Segmentation is a necessary first step for alignment. Moreover, by coregistering the different parameters of a multiparametric MRI examination, it becomes easier to develop automated decision support systems that automatically identify suspicious regions within the prostate. Segmentation is also a necessary step in fusing MRI to other imaging studies such as ultrasound (for biopsy or therapy), PET (for diagnosis), or future MRI studies (for longitudinal studies). Therefore, an automated segmentation tool of the prostate has broad appeal. In this study, we present a fully automated tool for prostate segmentation using multiplanar T2-weighted MRI and compare the results with manual tridimensional measurements using the ellipsoid formula, manual segmentation, and actual prostate volumes of prostatectomy specimens. Materials and Methods Study Design and Patient Population This retrospective single-institution study was approved by the local institutional review board and was compliant with the HIPAA. One hundred consecutive patients were enrolled in the study between June 2009 and October 2011. The mean age was 60 years (median, 60.6 years; range, 39–75 years) and the mean serum PSA level was 9.75 ng/dL (median, 6.85 ng/dL; range, 0.41–55.7 ng/dL). The inclusion criteria required that patients subsequently undergo robotic assisted radical prostatectomy. The patient population included 98 patients (the surgery was canceled for one patient, and another patient was excluded because his prostate gland was treated previously, which affects the signal characteristics of the gland) with a mean age of 60 years (median, 60.6 years; range, 39–74.5 years) and a mean serum PSA of 9.75 ng/dL (median, 6.85 ng/dL; range, 0.41–55.7 ng/dL).
MRI Technique All MRI studies were performed using an endorectal coil (BPX-30, Medrad) tuned to 127.8 MHz and a 16-channel cardiac coil (Sense, Philips Healthcare) on a 3-T magnet (Achieva, Philips
Healthcare) without prior bowel preparation. The endorectal coil was inserted using a semianesthetic gel (Lidocaine, Akorn) while the patient was in the left lateral decubitus position. The balloon surrounding the coil was distended with perfluorocarbon (3 mol/L [Fluorinert, 3M]) to a volume of approximately 45 mL to reduce susceptibility artifacts induced by air in the coil’s balloon. The protocol included triplanar T2-weighted turbo spin-echo (TSE) MRI, diffusion-weighted MRI, 3D MR spectroscopy, axial unenhanced T1-weighted MRI, and axial 3D fast-field echo dynamic contrast-enhanced MRI sequences. For this study, only triplane T2-weighted TSE MR images were used for volume determinations. The parameters of the T2-weighted images were as follows: FOV, 140 × 140 mm; acquisition matrix, 304 × 234; TR/TE, 8869/120; flip angle, 90°; slice thickness, 3 mm without gaps; image reconstruction, 512 × 512 pixels; and scanning resolution, 0.27 × 0.27 × 3.00 mm/pixel. The mean time interval between MRI and radical prostatectomy was 84 days (median, 69 days; range, 1–329 days).
Ellipsoid Prostate Volume Determination The greatest three dimensions of the prostate on MRI was measured manually and these measurements were used to determine the volume estimate of the prostate using the ellipsoid formula:
by appearance, shape, and topology information of the individual prostate subregions [19]. The segmentation outcome is presented by both 3D prostate surface rendering and a set of 2D prostate cross-sectional contours overlaid on the slices of the original scan as illustrated in Figure 2. The volume of the segmented prostate is automatically provided. The average total segmentation time for each patient is approximately 10 seconds, as shown in Figure S3. (Fig. S3, a supplemental video, can be viewed by clicking Supplemental at the top of this article and then clicking the figure number on the Supplemental page.)
Prostate Volume Assessment for Manual and Automated Segmentations To obtain a single prostate volume estimate for each case consisting of three different scans, we used a thresholded probability map (TPM) approach [20]. In this scheme, each of three segmentations are spatially combined together into a single probability map, where each of the segmentations has a vote of one third. Then this probability map is thresholded at 0.5, denoted henceforth as TPM 0.5 as shown in Figure 4.
Validation of Volume Estimates With Prostatectomy Specimens
where SI is the superoinferior dimension; AP, anteroposterior dimension; and RL, right-left dimension.
Ninety-eight patients underwent radical prostatectomy and prostatectomy specimens with seminal vesicles were weighed by a pathologist. The specimen mass was used as ground truth for our data, as reported previously [21].
Manual Prostate Segmentation
Statistical Analysis
Prostate boundaries were manually traced in three planes on T2-weighted MRI by a radiologist with 5 years of experience in prostate MRI. Regions of interest were drawn on each slice of each plane using software (Medical Image Processing, Analysis and Visualization [MIPAV], Center for Information Technology, National Institutes of Health).
The relationship between prostate volume estimates based on the ellipsoid formula, manual segmentation, automated segmentation, and prostatectomy specimen volumes (i.e., true prostate volume) were analyzed using the Pearson correlation coefficient and linear regression analysis methods. All analyses were conducted using statistics software (SAS version 6.0.1, SAS Institute). A separate analysis was conducted to identify the true prostate volume and MRI volume estimation errors (Appendix 1). Manually drawn contours of the prostate were compared with automatically generated segmentation using the Dice similarity coefficient [22]. The Dice similarity coefficient is used to quantify spatial agreement between manually and automatically annotated shapes. If 3D volumetric representations of manually and automatically found shapes are denoted as Vx and Vy, respectively, the Dice similarity coefficient (DSC) can be computed as follows:
(SI × AP × RL × 0.52),
Automated Prostate Segmentation The segmentation algorithm consists of two sequential steps: prostate localization and prostate contour refinement as shown in Figure 1. In the localization step, the MR image volume is examined for the regions that appear similar to a typical prostate and a mean prostate shape model is initialized at the most likely location. This approach is based on normalized gradient cross-correlation that is robust to MR image intensity inconsistencies and local high-intensity artifacts [18]. The purpose of the contour refinement step is to deform the initialized mean shape so that its surface becomes accurately aligned with the prostate boundary in the MR image data. The refinement uses a graph-search–based framework that performs the 3D deformation driven
DSC(Vx, Vy) =
Vx∩Vy Vx + Vy
.
AJR:201, November 2013 W721
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Turkbey et al. The value of the Dice similarity coefficient can vary between 0.0 (no overlap between the shapes) and 1.0 (perfect overlap); larger values correspond to better spatial agreement between the manually and automatically annotated shapes. Although the Dice similarity coefficient is a popular measure of segmentation accuracy, its major drawback is that manually drawn contours are inaccurate in the surface regions tangent to the image viewing plane—for example, the base and apex of the prostate on axial images. This inaccuracy contributes to large errors in the volume estimate. As a result, the manual contours may be incomplete in these regions, as shown in Figure 5B. To fairly compare an automatic segmentation with a set of manually drawn contours, we introduce the concept of a partial Dice similarity coefficient. In this measure, we exclude the portions in the automated segmentation that do not have corresponding manual contours. This concept is illustrated in Figure 5, where the portions of the prostate outside the dashed lines in Figure 5C are not considered when computing the partial Dice similarity coefficient.
positive correlation between true prostate volume and prostate volume estimates derived from the ellipsoid formula (R = 0.86– 0.90, p < 0.0001), manual segmentation (R =
0.89–0.91, p < 0.0001), and automated segmentation (R = 0.88–0.91, p < 0.0001) (Table 2). Moreover, there was a strong positive correlation between thresholded prostate vol-
TABLE 1: Estimated Volumes Derived From Manual and Fully Automated Segmentations, Ellipsoid Formula–Derived Volumes, and True Prostate Volume Extracted From Prostatectomy Specimen Prostate Volume (cm3)
Method of Estimating Prostate Volume and True Prostate Volume
Mean
SD
35.35
14.30
Axial
39.12
18.99
Coronal
37.36
17.86
Sagittal
35.78
17.74
Mean
37.42
18.08
TPM 0.7
28.92
15.57
TPM 0.5
37.81
18.15
TPM 0.3
45.15
20.64
51.67
18.10
Axial
39.39
19.01
Coronal
39.96
18.73
Sagittal
43.81
19.98
Mean
41.05
19.09
TPM 0.7
34.50
17.02
Ellipsoid volume Manual segmentation
True prostate volumea Fully automated segmentation
Results Table 1 shows the mean true prostate volume and prostate volume estimates obtained with the ellipsoid formula, manual segmentation, and automated segmentation. All prostate volume estimates (ellipsoid, manual, and automated) were smaller than the true prostate volume because the ground truth volume included the seminal vesicles whereas the segmented images did not (Fig. 6). A Pearson correlation analysis revealed a strong
TPM 0.5
40.12
19.01
TPM 0.3
48.29
21.38
Note—TPM = thresholded probability map. aExtracted from prostatectomy specimens.
TABLE 2: Pearson Correlation Coefficient Matrix for Prostate Volume Estimates Based on Ellipsoid Formula, Single-Plane Manual Segmentation, and Single-Plane Fully Automated Segmentation and for True Prostate Volumes Method of Estimating Prostate Volume and True Prostate Volume Ellipsoid volume
Manual Segmentation
Ellipsoid Volume
Axial
Coronal
Sagittal
—
0.90
0.90
0.89
Mean
True Prostate Volumea
Fully Automated Segmentation Axial
Coronal
Sagittal
Mean
0.90
0.86
0.91
0.92
0.91
0.92
Manual segmentation Axial
0.90
—
0.99
0.98
0.99
0.90
0.97
0.96
0.96
0.97
Coronal
0.90
0.99
—
0.98
0.99
0.90
0.97
0.96
0.95
0.97
Sagittal
0.89
0.98
0.98
—
0.99
0.89
0.97
0.95
0.94
0.96
Mean
0.90
0.99
0.99
0.99
—
0.90
0.98
0.97
0.96
0.97
0.86
0.90
0.90
0.89
0.90
—
0.90
0.90
0.88
0.90
Axial
0.91
0.97
0.97
0.97
0.98
0.90
—
0.98
0.97
0.99
True prostate volumea Fully automated segmentation Coronal
0.92
0.96
0.96
0.94
0.96
0.88
0.97
—
0.98
0.99
Sagittal
0.91
0.96
0.95
0.94
0.96
0.88
0.97
0.98
—
0.99
Mean
0.92
0.97
0.97
0.96
0.97
0.90
0.99
0.99
0.99
—
Note—Dash (—) indicates not applicable. aExtracted from prostatectomy specimens.
W722
AJR:201, November 2013
Automated Prostate Segmentation on MRI
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
TABLE 3: Pearson Correlation Coefficient Matrix for Thresholded Probability Map (TPM) Prostate Volume Estimates Derived From Ellipsoid Formula, Manual Segmentation, and Fully Automated Segmentation and for True Prostate Volumes Method of Estimating Prostate Volume and True Prostate Volume Ellipsoid volume
Manual Segmentation
Ellipsoid Volume
TPM 0.7
TPM 0.5
—
0.88
0.90
TPM 0.3
True Prostate Volumea
Fully Automated Segmentation TPM 0.7
TPM 0.5
TPM 0.3
0.90
0.86
0.91
0.92
0.92
Manual segmentation TPM 0.7
0.88
—
0.99
0.99
0.90
0.96
0.96
0.95
TPM 0.5
0.90
0.99
—
0.99
0.91
0.98
0.98
0.97
TPM 0.3
0.90
0.99
0.99
—
0.90
0.97
0.98
0.96
0.86
0.90
0.91
0.90
—
0.91
0.90
0.89
TPM 0.7
0.91
0.96
0.98
0.97
0.91
—
0.99
0.98
TPM 0.5
0.92
0.96
0.98
0.98
0.90
0.99
—
0.99
TPM 0.3
0.92
0.95
0.97
0.96
0.89
0.98
0.99
—
True prostate volumea Fully automated segmentation
Note—Dash (—) indicates not applicable. aExtracted from prostatectomy specimens.
ume estimates derived from manual and automated segmentations (Table 3). In linear regression analysis, there was a stronger correlation between true prostate volume and prostate volume estimates derived from manual and automated segmentations and between prostate volume estimates derived from manual and automated segmentations compared with the correlation between true prostate volume and the ellipsoid formula–derived volume (Fig. 7). Mean partial and full Dice similarity coefficients of 0.92 and 0.89 were achieved for axial automated segmentations, whereas the full Dice similarity coefficients obtained for TPM 0.3, 0.5, and 0.7 were 0.90, 0.85, and 0.89, respectively (Table 4 and Fig. 8). Discussion Prostate MRI is increasingly used in the management of patients with known or suspected prostate cancer. Prostate volume determination can be important for diagnosis, prognosis, biopsy planning, and treatment decisions. Measurements based on the DRE are subjective and difficult to reproduce. Prostate volumes determined by the ellipsoid formula correlate with actual prostate volumes surprisingly well; however, the other benefits of segmentation—namely, the ability to coregister other modalities and perform more advanced imaging processing—are not possible with simple trilinear measurements. Previous studies have reported various segmentation methods of prostate MR images. Jia et al. [23] used a semiautomated prostate segmentation
tool that included planimetry of eight slices in the axial and coronal planes in 24 beagles and reported strong correlation (R = 0.98) of segmented volumes and necropsy specimens. Pasquier et al. [24] assessed a 3D deformable model and a region-growing algorithm for automatic delineation of the prostate and reported a volume ratio (automated/manual estimated volume) of 1.13 in 24 patients. Klein et al. [25] evaluated an automatic segmentation method using atlas matching based on localized mutual information in 50 patients; they reported a median Dice similarity coefficient of 0.85 and segmentation errors of 1 and 1.5 mm in 50% and 75% of patients, respectively. Martin et al. [26] used an automated segmentation approach based on a probabilistic atlas and a spatially constrained deformable model in 36 patients and reported a median Dice similarity coefficient of 0.86. Recently, Bulman et al. [27] compared prostate volumes obtained by automated computer-generated multifeature active shape models based on MRI with pathology specimen volumes in 91 patients and reported that multifeature active shape models had the highest slope (0.888) with a concordance correlation coefficient of 0.867 with respect to prostatectomy specimen volume. However, this system requires the operator to identify the center of the prostate on a single midgland axial T2-weighted section; thus, it is not completely automated. The method used in our study differs from prior approaches in that it includes a 3D approach and uses normalized gradient fields cross-correlation and a graph-based search. It
yielded reasonable and encouraging correlation with both true prostate volume of the specimens and with manual segmentation volumes in a fully automatic and highly time-efficient (≈ 10 seconds) manner without the need for cursor placement by the user. Thus, this method avoids subjective differences among different viewers and yields a highly reproducible result. The best prostate volume estimates obtained with manual segmentation were quite close to those obtained with the fully automated system; for example, the Pearson correlation coefficient between manual and automated axial prostate volume estimates and true prostate volume (from prostatectomy specimen) were 0.90, whereas for manual and automated TPM 0.5 prostate volume were 0.91 and 0.90, respectively. The partial mean Dice similarity coefficient (which excludes slices that were not annotated) for triplane automated segmentations ranged between 0.90 and 0.92, whereas the full mean Dice similarity coefficient (which includes all slices) ranged between 0.83 and 0.89, with 0.89 representing the axial full mean Dice similarity coefficient. Thus, this automated prostate segmentation tool can provide a convenient way to estimate prostate volume and to segment the prostate, which can potentially be used in clinical management of prostate cancer patients and in research protocols. Our study has several limitations. First, we used fresh prostatectomy specimen weight as the ground truth to validate our results and the mean specimen volume was higher than either
AJR:201, November 2013 W723
Turkbey et al. TABLE 4: Partial and Full Dice Similarity Coefficient Values Between Manually and Fully Automated Segmented Single-Plane and Thresholded Probability Map (TPM) Shapes Dice Similarity Coefficient Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Shapes
Mean
SD
Standard Error
Partial Dice similarity coefficient Axial
0.92
0.03
0.003
Sagittal
0.90
0.03
0.003
Coronal
0.91
0.03
0.003
Axial
0.89
0.04
0.004
Sagittal
0.87
0.04
0.004
Coronal
0.83
0.06
0.006
TPM 0.7
0.89
0.04
0.004
Full Dice similarity coefficient
TPM 0.5
0.85
0.05
0.005
TPM 0.3
0.90
0.03
0.003
the manual or automated volume estimates because the seminal vesicles were not amputated from the specimen before weighing. Moreover, variable amounts of extraprostatic tissue may be included in some specimens when the surgeon decided to perform a wider resection around the prostate. There was no reliable way to correct for this discrepancy; however, we believe that the seminal vesicles and minimal periprostatic fat contribute modestly to the weight of the prostate. It is also likely, however, that the ex vivo specimen is somewhat smaller because of the loss of blood from the gland. This problem in comparisons with ex vivo tissue is unavoidable. An additional issue is that the manual segmentations were performed by a single experienced operator. However, this reader’s results correlated well with the histopathology specimen and we believe that those results are reliable. Third, the MR images were obtained with an endorectal coil, which compresses the gland posteriorly, potentially affecting the volume estimation of the prostate compared with specimens [28]. We believe that this effect is small and the value of the endorectal coil is that it provides images with superior resolution for delineating the prostate boundaries, which is critical for accurate volume determinations. Anecdotally, this automated segmentation system performs well on images obtained without an endorectal coil. Thus, on balance, this limitation probably contributes minimally to disparities between pathologic specimens and MRI-derived volumes. In conclusion, 3D segmentations of the prostate and prostate volumes derived from using an automated segmentation tool based on nor-
W724
malized gradient fields, cross-correlation, and graph-search refinement can yield accurate prostate volume determinations in a remarkably time-efficient manner. This time-consuming task can potentially be used routinely because it requires essentially no user input and only 10 seconds to complete. Applications of accurate prostate segmentation go beyond simple volume determinations, PSA density measurements, and follow-up of benign prostate hyperplasia and extend to multimodal image fusion and multi–time point image fusion with implications for automated detection, biopsy, and imaging-guided therapy. References 1. Ramon J, Boccon-Gibod L, Billebaud T, et al. Prostate-specific antigen density: a means to enhance detection of prostate cancer. Eur Urol 1994; 25:288–294 2. Blackwell KL, Bostwick DG, Myers RP, Zincke H, Oesterling JE. Combining prostate specific antigen with cancer and gland volume to predict more reliably pathological stage: the influence of prostate specific antigen cancer density. J Urol 1994; 151:1565–1570 3. Stephan C, Cammann H, Meyer HA, Lein M, Jung K. PSA and new biomarkers within multivariate models to improve early detection of prostate cancer. Cancer Lett 2007; 249:18–29 4. Corcoran NM, Casey RG, Hong MK, et al. The ability of prostate-specific antigen (PSA) density to predict an upgrade in Gleason score between initial prostate biopsy and prostatectomy diminishes with increasing tumour grade due to reduced PSA secretion per unit tumour volume. BJU Int 2012; 110:36–42
5. Choi SY, Chang IH, Kim YS, Kim TH, Kim W, Myung SC. Prostate specific antigen velocity per prostate volume: a novel tool for prostate biopsy prediction. Urology 2011; 78:874–879 6. Ko JS, Landis P, Carter HB, Partin AW. Effect of intra-observer variation in prostate volume measurement on prostate-specific antigen density calculations among prostate cancer active surveillance participants. BJU Int 2011; 108:1739–1742 7. Sfoungaristos S, Perimenis P. PSA density versus risk stratification for lymphadenectomy-making decision in patients with prostate cancer undergoing radical prostatectomy. Int Urol Nephrol 2011; 43:1073–1079 8. Eskicorapci SY, Guliyev F, Akdogan B, Dogan HS, Ergen A, Ozen H. Individualization of the biopsy protocol according to the prostate gland volume for prostate cancer detection. J Urol 2005; 173:1536–1540 9. Al-Qaisieh B, Ash D, Bottomley DM, Carey BM. Impact of prostate volume evaluation by different observers on CT-based post-implant dosimetry. Radiother Oncol 2002; 62:267–273 10. Vikal S, Haker S, Tempany C, Fichtinger G. Prostate contouring in MRI guided biopsy. Proc SPIE 2009; 7259:72594A 11. Giubilei G, Ponchietti R, Biscioni S, et al. Accuracy of prostate volume measurements using transrectal multiplanar three-dimensional sonography. Int J Urol 2005; 12:936–938 12. Matthews GJ, Motta J, Fracehia JA. The accuracy of transrectal ultrasound prostate volume estimation: clinical correlations. J Clin Ultrasound 1996; 24:501–505 13. Rahmouni A, Yang A, Tempany CM, et al. Accuracy of in-vivo assessment of prostatic volume by MRI and transrectal ultrasonography. J Comput Assist Tomogr 1992; 16:935–940 14. Sosna J, Rofsky NM, Gaston SM, DeWolf WC, Lenkinski RE. Determinations of prostate volume at 3-Tesla using an external phased array coil: comparison to pathologic specimens. Acad Radiol 2003; 10:846–853 15. Smith WL, Lewis C, Bauman G, et al. Prostate volume contouring: a 3D analysis of segmentation using 3DTRUS, CT, and MR. Int J Radiat Oncol Biol Phys 2007; 67:1238–1247 16. Lee JS, Chung BH. Transrectal ultrasound versus magnetic resonance imaging in the estimation of prostate volume as compared with radical prostatectomy specimens. Urol Int 2007; 78:323–327 17. Jeong CW, Park HK, Hong SK, Byun SS, Lee HJ, Lee SE. Comparison of prostate volume measured by transrectal ultrasonography and MRI with the actual prostate volume measured after radical prostatectomy. Urol Int 2008; 81:179–185 18. Fotin SV, Yin Y, Periaswamy S, et al. Normalized gradient fields cross-correlation for automated de-
AJR:201, November 2013
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI tection of prostate in magnetic resonance images. Proc SPIE 2012; 8314:83140V: 19. Yin Y, Fotin SV, Periaswamy S, et al. Fully automated prostate segmentation in 3D MR based on normalized gradient field cross-correlation initialization and LOGISMOS refinement. Proc SPIE 2012; 8314:831406 20. Biancardi AM, Jirapatnakul AC, Reeves AP. A comparison of ground truth estimation methods. Int J Comput Assist Radiol Surg 2010; 5:295–305 21. Rodriguez E Jr, Skarecky D, Narula N, Ahlering TE. Prostate volume estimation using the ellipsoid formula consistently underestimates actual gland size. J Urol 2008; 179:501–503 22. Dice LR. Measures of the amount of ecologic asso-
ciation between species. Ecology 1945; 26:297–302 23. Jia G, Baudendistel KT, von Tengg-Kobligk H, et al. Assessing prostate volume by magnetic resonance imaging: a comparison of different measurement approaches for organ volume analysis. Invest Radiol 2005; 40:243–248 24. Pasquier D, Lacornerie T, Vermandel M, Rousseau J, Lartigau E, Betrouni N. Automatic segmentation of pelvic structures from magnetic resonance images for prostate cancer radiotherapy. Int J Radiat Oncol Biol Phys 2007; 68:592–600 25. Klein S, van der Heide UA, Lips IM, van Vulpen M, Staring M, Pluim JP. Automatic segmentation of the prostate in 3D MR images by atlas matching using localized mutual information. Med Phys
2008; 35:1407–1417 26. Martin S, Troccaz J, Daanenc V. Automated segmentation of the prostate in 3D MR images using a probabilistic atlas and a spatially constrained deformable model. Med Phys 2010; 37:1579–1590 27. Bulman JC, Toth R, Patel AD, et al. Automated computer-derived prostate volumes from MR imaging data: comparison with radiologist-derived MR imaging and pathologic specimen volumes. Radiology 2012; 262:144–151 28. Heijmink SW, Scheenen TW, van Lin EN, et al. Changes in prostate shape and volume and their implications for radiotherapy after introduction of endorectal balloon as determined by MRI at 3T. Int J Radiat Oncol Biol Phys 2009; 73:1446–1453
APPENDIX 1: Mass and Volume Estimation Errors In another experiment we estimated how accurately we can predict true prostate mass from the prostate volume obtained using MR images. The quality of such an estimate depends on several factors such as the accuracy of the segmentation, geometric distortion caused by MRI, and minor variations in the density of the prostate tissue. The relative error in the prostate mass estimate is the upper bound of the error in physical volume estimate because the density variation does not have an impact on the volume. To measure the mass estimation error, we split available studies into training and test subsets. From the training subset, we constructed two models: The objective of the first model was to predict the true mass m using the volume Vm from manually traced axial contours and then using the volume from the automated segmentation (Va) obtained with thresholded probability map (TPM) 0.5 method. For modeling we used linear regression in the following form: m = α + V + β, where α is a multiplicative component that roughly represents prostate tissue density and β is an additive component to account for variation in measured mass due to the seminal vesicles and the excess of the extracted tissue.
Localization
Linear regression parameters were estimated using least squares over the training set and the following coefficients were established for MRI-derived segmentations: α = 0.820 g/cm3 and β = 18.680 g for manual segmentation and α = 0.804 g/cm3 and β = 17.830 g for automated segmentation. These linear regression models were then applied to each of the studies in the test subset and a mass estimation error was calculated from measured prostate mass m and estimated mass me. The absolute error (Δm = me – m) and relative error (δm = Δm / m) were calculated for each of the studies. For manually obtained segmentation, the absolute and relative errors were in the range of from –25.92 to 11.39 g and from –30.03% to 30.13%, respectively. For 95% of the studies, the estimated mass fell within 28.25% of the measured mass. The root mean squared error for manual segmentation was 13.90%. For automatically obtained segmentation, the absolute and relative errors were in the range of from –20.45 to 9.76 g and from –32.26% to 31.38%, respectively. Here for 95% of the studies the estimated mass fell within 25.73% of the measured mass. The root mean squared error for automatic segmentation was 13.10%. Thus, MR images can also be used to effectively estimate the prostate mass.
Refinement
Fig. 1—Automated prostate segmentation consists of two steps: localization and refinement. Red shows slice of 3D prostate bounding box as identified by localization step, and green shows evolution of prostate surface during refinement step.
AJR:201, November 2013 W725
Turkbey et al.
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Fig. 2—Prostate segmentation visualization modes. A, Three-dimensional surface rendering shows prostate (green). B, Two-dimensional cross sections were overlaid on original MR image slice to produce this image. Prostate is outlined in green.
A
B Fig. S3—Screen shot from video of fully automated segmentation tool (Medical Image Processing, Analysis and Visualization [MIPAV], Center for Information Technology, National Institutes of Health) in sample case. (To view this video, click Supplemental at the top of this article and then click the figure number on the Supplemental page.)
A
W726
B
Fig. 4—Thresholded probability map. A, Sketch shows original three segmentations. B, Sketch shows probability map thresholded at level of 0.5.
AJR:201, November 2013
A
B
C
Fig. 5—Partial Dice similarity coefficient. A, Axial MR image with manually drawn contour showing prostate (green). B, Sagittal MR image shows cross sections (green lines) of multiple manually drawn axial contours. C, Sagittal MR image shows portions of prostate removed (dashed lines) for calculation of partial Dice similarity coefficient. Green shows outline of prostate.
80 70
Manual segmentation Automated segmentation True prostate volume
80
60
70
50
Mean Prostate Volume (cm3)
Mean Prostate Volume (cm3)
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI
40 30 20 10 0
Axial
Coronal
Sagittal
Manual segmentation Automated segmentation True prostate volume
60 50 40 30 20 10
Mean of Single-Plane Values
0
TPM 0.7
TPM 0.5
TPM 0.3
A
B
Fig. 6—Bar graphs show mean prostate volume estimates and mean true prostate volume. Vertical lines show mean ± 1 SD. A, Bar graph shows prostate volume estimates based on manual segmentation and automated segmentation and true prostate volumes extracted from prostatectomy specimens in three imaging planes and mean values overall. B, Bar graph shows prostate volume estimates based on thresholded probability map (TPM) approaches at 0.7, 0.5, and 0.3 and true prostate volumes extracted from prostatectomy specimens.
AJR:201, November 2013 W727
D
G
180 160 140 120 100 80 60 40 20 0
150 50 100 PV in Axial Plane Using Automated Segmentation (cm3)
150 50 100 PV in Axial Plane Using Manual 3 Segmentation (cm )
20
B 160 140 120 100 80 60 40 20 0
200
0
200
H
180 160 140 120 100 80 60 40 20 0
40 60 80 100 120 140 160 Mean PV Using Manual Segmentation (cm3)
50 100 150 Mean PV Using Automated Segmentation (cm3)
20
40 60 80 100 120 140 160 Mean PV Using Manual Segmentation (cm3)
y = 0.8948x + 17.857 R 2 = 0.8211
C 160 140 120 100 80 60 40 20 0
200
y = 1.0276x + 2.5995 R 2 = 0.9475
0
160 140 120 100 80 60 40 20 0 0
y = 0.8459x + 16.963 R 2 = 0.81
E
y = 0.9744x + 1.2638 R 2 = 0.9468
0
True PV (cm3) 0
y = 0.852x + 18.116 R 2 = 0.8163
0
y = 0.8961x + 18.16 R 2 = 0.8172
True PV (cm3)
160 140 120 100 80 60 40 20 0
200
True PV (cm3)
True PV (cm3)
A
50 100 150 PV in Axial Plane Using Manual 3 Segmentation (cm )
160 140 120 100 80 60 40 20 0
F
I
180 160 140 120 100 80 60 40 20 0
20 40 60 80 100 120 140 160 PV Using Manual Segmentation With TPM 0.5 (cm3)
y = 0.8464x + 17.724 R 2 = 0.8044
0
PV Using Automated Segmentation With TPM 0.5 (cm3)
True PV (cm3)
y = 0.8541x + 18.291 R 2 = 0.8183
Mean PV Using Automated Segmentation (cm3)
True PV (cm3)
160 140 120 100 80 60 40 20 0 0
PV in Axial Plane Using Automated Segmentation (cm3)
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Turkbey et al.
100 150 50 PV Using Automated Segmentation With TPM 0.5 (cm3)
200
y = 1.0231x + 1.439 R 2 = 0.9548
0
20 40 60 80 100 120 140 160 PV Using Manual Segmentation With TPM 0.5 (cm3)
Fig. 7— Linear regression plots. A–I, Linear regression plots show prostate volume (PV) estimates derived from manual and automated segmentations and true prostate volumes extracted from prostatectomy specimens. TPM = thresholded probability map.
W728
AJR:201, November 2013
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI
A
B
C
D
E
F
Fig. 8—Sample segmentations. A–F, Axial (A and B), sagittal (C and D), and coronal (E and F) MR images show manual tracings (green) and automatically generated segmentations (red) of prostate. A, C, and E are images of 62-year-old man and B, D, and F are images of 56-year-old man.
F O R YO U R I N F O R M AT I O N
The data supplement accompanying this web exclusive article can be viewed by clicking “Supplemental” at the top of the article.
AJR:201, November 2013 W729
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Medical Physics and Informatics Original Research
Fully Automated Prostate Segmentation on MRI: Comparison With Manual Segmentation Methods and Specimen Volumes Baris Turkbey 1 Sergei V. Fotin2 Robert J. Huang1 Yin Yin2 Dagane Daar 1 Omer Aras1 Marcelino Bernardo1 Brian E. Garvey 1 Juanita Weaver 1 Hrishikesh Haldankar 2 Naira Muradyan2 Maria J. Merino 3 Peter A. Pinto 4 Senthil Periaswamy 2 Peter L. Choyke1 Turkbey B, Fotin SV, Huang RJ, et al. Keywords: fully automated segmentation, MRI, prostate, segmentation, volume DOI:10.2214/AJR.12.9712 Received July 21, 2012; accepted after revision September 26, 2012. S. Periaswamy is a shareholder in iCAD. 1
Molecular Imaging Program, National Cancer Institute, National Institutes of Health, 10 Center Dr, MSC 1182, Bldg 10, Rm B69, Bethesda, MD 20892-1088. Address correspondence to P. L. Choyke ([email protected]).
2
iCAD Inc., Nashua, NH.
3
Laboratory of Pathology, NCI, NIH, Bethesda, MD.
4
Urologic Oncology Branch, NCI, NIH, Bethesda, MD.
Supplemental Data Available online at www.ajronline.org. WEB This is a web exclusive article. AJR 2013; 201:W720–W729 0361–803X/13/2015–W720 © American Roentgen Ray Society
W720
OBJECTIVE. The objective of our study was to compare calculated prostate volumes derived from tridimensional MR measurements (ellipsoid formula), manual segmentation, and a fully automated segmentation system as validated by actual prostatectomy specimens. MATERIALS AND METHODS. Ninety-eight consecutive patients (median age, 60.6 years; median prostate-specific antigen [PSA] value, 6.85 ng/mL) underwent triplane T2weighted MRI on a 3-T magnet with an endorectal coil while undergoing diagnostic workup for prostate cancer. Prostate volume estimates were determined using the formula for ellipsoid volume based on tridimensional measurements, manual segmentation of triplane MRI, and automated segmentation based on normalized gradient fields cross-correlation and graph-search refinement. Estimates of prostate volume based on ellipsoid volume, manual segmentation, and automated segmentation were compared with prostatectomy specimen volumes. Prostate volume estimates were compared using the Pearson correlation coefficient and linear regression analysis. The Dice similarity coefficient was used to quantify spatial agreement between manual segmentation and automated segmentation. RESULTS. The Pearson correlation coefficient revealed strong positive correlation between prostatectomy specimen volume and prostate volume estimates derived from manual segmentation (R = 0.89–0.91, p < 0.0001) and automated segmentation (R = 0.88–0.91, p < 0.0001). No difference was observed between manual segmentation and automated segmentation. Mean partial and full Dice similarity coefficients of 0.92 and 0.89, respectively, were achieved for axial automated segmentation. CONCLUSION. Prostate volume estimates obtained with a fully automated 3D segmentation tool based on normalized gradient fields cross-correlation and graph-search refinement can yield highly accurate prostate volume estimates in a clinically relevant time of 10 seconds. This tool will assist in developing a broad range of applications including routine prostate volume estimations, image registration, biopsy guidance, and decision support systems.
S
egmentation of the prostate from surrounding tissue on MRI is useful for a variety of clinical purposes including determination of prostate volume, prostate-specific antigen (PSA) density, registration of MRI with other modalities such as ultrasound and PET, and imagingguided biopsy and therapy. Prostate volumes are correlated with lower urinary tract symptoms and are also relevant to decisions regarding the feasibility of brachytherapy and surgery [1–10]. The current methods of estimating prostate volume include physician estimation based on a digital rectal examination (DRE) and imaging measurements based on ultrasound, CT, or MRI. The imaging measurements generally rely on triplanar linear measurements to calculate prostate volume
using the following formula for the volume of an ellipsoid: (length × width × height × 0.52), which assumes that the prostate has a regular ovoid shape [11]. The DRE is often inaccurate because it is based on a subjective estimate of the examiner. Prostate volume determinations based on the ellipsoid formula are often inaccurate because the shape of the prostate varies dramatically [12]. As benign prostatic hyperplasia develops, the prostate evolves from a coneshaped organ to a more spheric organ that often includes an eccentrically enlarged median lobe that is not accounted for by the ellipsoid formula. Manual planimetry based on imaging sections is more accurate but is time-con-
AJR:201, November 2013
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI suming and requires expertise [13–17]. Automating this process could result in highly accurate yet timely prostate volume determinations that could be incorporated routinely into clinical interpretations. In addition to rapid prostate volume determinations, there are other reasons to justify automated segmentation systems. For multiparametric MRI, it is important to align all images (T2, diffusion-weighted, dynamic contrast-enhanced images) to offset misregistration due to patient motion. Segmentation is a necessary first step for alignment. Moreover, by coregistering the different parameters of a multiparametric MRI examination, it becomes easier to develop automated decision support systems that automatically identify suspicious regions within the prostate. Segmentation is also a necessary step in fusing MRI to other imaging studies such as ultrasound (for biopsy or therapy), PET (for diagnosis), or future MRI studies (for longitudinal studies). Therefore, an automated segmentation tool of the prostate has broad appeal. In this study, we present a fully automated tool for prostate segmentation using multiplanar T2-weighted MRI and compare the results with manual tridimensional measurements using the ellipsoid formula, manual segmentation, and actual prostate volumes of prostatectomy specimens. Materials and Methods Study Design and Patient Population This retrospective single-institution study was approved by the local institutional review board and was compliant with the HIPAA. One hundred consecutive patients were enrolled in the study between June 2009 and October 2011. The mean age was 60 years (median, 60.6 years; range, 39–75 years) and the mean serum PSA level was 9.75 ng/dL (median, 6.85 ng/dL; range, 0.41–55.7 ng/dL). The inclusion criteria required that patients subsequently undergo robotic assisted radical prostatectomy. The patient population included 98 patients (the surgery was canceled for one patient, and another patient was excluded because his prostate gland was treated previously, which affects the signal characteristics of the gland) with a mean age of 60 years (median, 60.6 years; range, 39–74.5 years) and a mean serum PSA of 9.75 ng/dL (median, 6.85 ng/dL; range, 0.41–55.7 ng/dL).
MRI Technique All MRI studies were performed using an endorectal coil (BPX-30, Medrad) tuned to 127.8 MHz and a 16-channel cardiac coil (Sense, Philips Healthcare) on a 3-T magnet (Achieva, Philips
Healthcare) without prior bowel preparation. The endorectal coil was inserted using a semianesthetic gel (Lidocaine, Akorn) while the patient was in the left lateral decubitus position. The balloon surrounding the coil was distended with perfluorocarbon (3 mol/L [Fluorinert, 3M]) to a volume of approximately 45 mL to reduce susceptibility artifacts induced by air in the coil’s balloon. The protocol included triplanar T2-weighted turbo spin-echo (TSE) MRI, diffusion-weighted MRI, 3D MR spectroscopy, axial unenhanced T1-weighted MRI, and axial 3D fast-field echo dynamic contrast-enhanced MRI sequences. For this study, only triplane T2-weighted TSE MR images were used for volume determinations. The parameters of the T2-weighted images were as follows: FOV, 140 × 140 mm; acquisition matrix, 304 × 234; TR/TE, 8869/120; flip angle, 90°; slice thickness, 3 mm without gaps; image reconstruction, 512 × 512 pixels; and scanning resolution, 0.27 × 0.27 × 3.00 mm/pixel. The mean time interval between MRI and radical prostatectomy was 84 days (median, 69 days; range, 1–329 days).
Ellipsoid Prostate Volume Determination The greatest three dimensions of the prostate on MRI was measured manually and these measurements were used to determine the volume estimate of the prostate using the ellipsoid formula:
by appearance, shape, and topology information of the individual prostate subregions [19]. The segmentation outcome is presented by both 3D prostate surface rendering and a set of 2D prostate cross-sectional contours overlaid on the slices of the original scan as illustrated in Figure 2. The volume of the segmented prostate is automatically provided. The average total segmentation time for each patient is approximately 10 seconds, as shown in Figure S3. (Fig. S3, a supplemental video, can be viewed by clicking Supplemental at the top of this article and then clicking the figure number on the Supplemental page.)
Prostate Volume Assessment for Manual and Automated Segmentations To obtain a single prostate volume estimate for each case consisting of three different scans, we used a thresholded probability map (TPM) approach [20]. In this scheme, each of three segmentations are spatially combined together into a single probability map, where each of the segmentations has a vote of one third. Then this probability map is thresholded at 0.5, denoted henceforth as TPM 0.5 as shown in Figure 4.
Validation of Volume Estimates With Prostatectomy Specimens
where SI is the superoinferior dimension; AP, anteroposterior dimension; and RL, right-left dimension.
Ninety-eight patients underwent radical prostatectomy and prostatectomy specimens with seminal vesicles were weighed by a pathologist. The specimen mass was used as ground truth for our data, as reported previously [21].
Manual Prostate Segmentation
Statistical Analysis
Prostate boundaries were manually traced in three planes on T2-weighted MRI by a radiologist with 5 years of experience in prostate MRI. Regions of interest were drawn on each slice of each plane using software (Medical Image Processing, Analysis and Visualization [MIPAV], Center for Information Technology, National Institutes of Health).
The relationship between prostate volume estimates based on the ellipsoid formula, manual segmentation, automated segmentation, and prostatectomy specimen volumes (i.e., true prostate volume) were analyzed using the Pearson correlation coefficient and linear regression analysis methods. All analyses were conducted using statistics software (SAS version 6.0.1, SAS Institute). A separate analysis was conducted to identify the true prostate volume and MRI volume estimation errors (Appendix 1). Manually drawn contours of the prostate were compared with automatically generated segmentation using the Dice similarity coefficient [22]. The Dice similarity coefficient is used to quantify spatial agreement between manually and automatically annotated shapes. If 3D volumetric representations of manually and automatically found shapes are denoted as Vx and Vy, respectively, the Dice similarity coefficient (DSC) can be computed as follows:
(SI × AP × RL × 0.52),
Automated Prostate Segmentation The segmentation algorithm consists of two sequential steps: prostate localization and prostate contour refinement as shown in Figure 1. In the localization step, the MR image volume is examined for the regions that appear similar to a typical prostate and a mean prostate shape model is initialized at the most likely location. This approach is based on normalized gradient cross-correlation that is robust to MR image intensity inconsistencies and local high-intensity artifacts [18]. The purpose of the contour refinement step is to deform the initialized mean shape so that its surface becomes accurately aligned with the prostate boundary in the MR image data. The refinement uses a graph-search–based framework that performs the 3D deformation driven
DSC(Vx, Vy) =
Vx∩Vy Vx + Vy
.
AJR:201, November 2013 W721
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Turkbey et al. The value of the Dice similarity coefficient can vary between 0.0 (no overlap between the shapes) and 1.0 (perfect overlap); larger values correspond to better spatial agreement between the manually and automatically annotated shapes. Although the Dice similarity coefficient is a popular measure of segmentation accuracy, its major drawback is that manually drawn contours are inaccurate in the surface regions tangent to the image viewing plane—for example, the base and apex of the prostate on axial images. This inaccuracy contributes to large errors in the volume estimate. As a result, the manual contours may be incomplete in these regions, as shown in Figure 5B. To fairly compare an automatic segmentation with a set of manually drawn contours, we introduce the concept of a partial Dice similarity coefficient. In this measure, we exclude the portions in the automated segmentation that do not have corresponding manual contours. This concept is illustrated in Figure 5, where the portions of the prostate outside the dashed lines in Figure 5C are not considered when computing the partial Dice similarity coefficient.
positive correlation between true prostate volume and prostate volume estimates derived from the ellipsoid formula (R = 0.86– 0.90, p < 0.0001), manual segmentation (R =
0.89–0.91, p < 0.0001), and automated segmentation (R = 0.88–0.91, p < 0.0001) (Table 2). Moreover, there was a strong positive correlation between thresholded prostate vol-
TABLE 1: Estimated Volumes Derived From Manual and Fully Automated Segmentations, Ellipsoid Formula–Derived Volumes, and True Prostate Volume Extracted From Prostatectomy Specimen Prostate Volume (cm3)
Method of Estimating Prostate Volume and True Prostate Volume
Mean
SD
35.35
14.30
Axial
39.12
18.99
Coronal
37.36
17.86
Sagittal
35.78
17.74
Mean
37.42
18.08
TPM 0.7
28.92
15.57
TPM 0.5
37.81
18.15
TPM 0.3
45.15
20.64
51.67
18.10
Axial
39.39
19.01
Coronal
39.96
18.73
Sagittal
43.81
19.98
Mean
41.05
19.09
TPM 0.7
34.50
17.02
Ellipsoid volume Manual segmentation
True prostate volumea Fully automated segmentation
Results Table 1 shows the mean true prostate volume and prostate volume estimates obtained with the ellipsoid formula, manual segmentation, and automated segmentation. All prostate volume estimates (ellipsoid, manual, and automated) were smaller than the true prostate volume because the ground truth volume included the seminal vesicles whereas the segmented images did not (Fig. 6). A Pearson correlation analysis revealed a strong
TPM 0.5
40.12
19.01
TPM 0.3
48.29
21.38
Note—TPM = thresholded probability map. aExtracted from prostatectomy specimens.
TABLE 2: Pearson Correlation Coefficient Matrix for Prostate Volume Estimates Based on Ellipsoid Formula, Single-Plane Manual Segmentation, and Single-Plane Fully Automated Segmentation and for True Prostate Volumes Method of Estimating Prostate Volume and True Prostate Volume Ellipsoid volume
Manual Segmentation
Ellipsoid Volume
Axial
Coronal
Sagittal
—
0.90
0.90
0.89
Mean
True Prostate Volumea
Fully Automated Segmentation Axial
Coronal
Sagittal
Mean
0.90
0.86
0.91
0.92
0.91
0.92
Manual segmentation Axial
0.90
—
0.99
0.98
0.99
0.90
0.97
0.96
0.96
0.97
Coronal
0.90
0.99
—
0.98
0.99
0.90
0.97
0.96
0.95
0.97
Sagittal
0.89
0.98
0.98
—
0.99
0.89
0.97
0.95
0.94
0.96
Mean
0.90
0.99
0.99
0.99
—
0.90
0.98
0.97
0.96
0.97
0.86
0.90
0.90
0.89
0.90
—
0.90
0.90
0.88
0.90
Axial
0.91
0.97
0.97
0.97
0.98
0.90
—
0.98
0.97
0.99
True prostate volumea Fully automated segmentation Coronal
0.92
0.96
0.96
0.94
0.96
0.88
0.97
—
0.98
0.99
Sagittal
0.91
0.96
0.95
0.94
0.96
0.88
0.97
0.98
—
0.99
Mean
0.92
0.97
0.97
0.96
0.97
0.90
0.99
0.99
0.99
—
Note—Dash (—) indicates not applicable. aExtracted from prostatectomy specimens.
W722
AJR:201, November 2013
Automated Prostate Segmentation on MRI
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
TABLE 3: Pearson Correlation Coefficient Matrix for Thresholded Probability Map (TPM) Prostate Volume Estimates Derived From Ellipsoid Formula, Manual Segmentation, and Fully Automated Segmentation and for True Prostate Volumes Method of Estimating Prostate Volume and True Prostate Volume Ellipsoid volume
Manual Segmentation
Ellipsoid Volume
TPM 0.7
TPM 0.5
—
0.88
0.90
TPM 0.3
True Prostate Volumea
Fully Automated Segmentation TPM 0.7
TPM 0.5
TPM 0.3
0.90
0.86
0.91
0.92
0.92
Manual segmentation TPM 0.7
0.88
—
0.99
0.99
0.90
0.96
0.96
0.95
TPM 0.5
0.90
0.99
—
0.99
0.91
0.98
0.98
0.97
TPM 0.3
0.90
0.99
0.99
—
0.90
0.97
0.98
0.96
0.86
0.90
0.91
0.90
—
0.91
0.90
0.89
TPM 0.7
0.91
0.96
0.98
0.97
0.91
—
0.99
0.98
TPM 0.5
0.92
0.96
0.98
0.98
0.90
0.99
—
0.99
TPM 0.3
0.92
0.95
0.97
0.96
0.89
0.98
0.99
—
True prostate volumea Fully automated segmentation
Note—Dash (—) indicates not applicable. aExtracted from prostatectomy specimens.
ume estimates derived from manual and automated segmentations (Table 3). In linear regression analysis, there was a stronger correlation between true prostate volume and prostate volume estimates derived from manual and automated segmentations and between prostate volume estimates derived from manual and automated segmentations compared with the correlation between true prostate volume and the ellipsoid formula–derived volume (Fig. 7). Mean partial and full Dice similarity coefficients of 0.92 and 0.89 were achieved for axial automated segmentations, whereas the full Dice similarity coefficients obtained for TPM 0.3, 0.5, and 0.7 were 0.90, 0.85, and 0.89, respectively (Table 4 and Fig. 8). Discussion Prostate MRI is increasingly used in the management of patients with known or suspected prostate cancer. Prostate volume determination can be important for diagnosis, prognosis, biopsy planning, and treatment decisions. Measurements based on the DRE are subjective and difficult to reproduce. Prostate volumes determined by the ellipsoid formula correlate with actual prostate volumes surprisingly well; however, the other benefits of segmentation—namely, the ability to coregister other modalities and perform more advanced imaging processing—are not possible with simple trilinear measurements. Previous studies have reported various segmentation methods of prostate MR images. Jia et al. [23] used a semiautomated prostate segmentation
tool that included planimetry of eight slices in the axial and coronal planes in 24 beagles and reported strong correlation (R = 0.98) of segmented volumes and necropsy specimens. Pasquier et al. [24] assessed a 3D deformable model and a region-growing algorithm for automatic delineation of the prostate and reported a volume ratio (automated/manual estimated volume) of 1.13 in 24 patients. Klein et al. [25] evaluated an automatic segmentation method using atlas matching based on localized mutual information in 50 patients; they reported a median Dice similarity coefficient of 0.85 and segmentation errors of 1 and 1.5 mm in 50% and 75% of patients, respectively. Martin et al. [26] used an automated segmentation approach based on a probabilistic atlas and a spatially constrained deformable model in 36 patients and reported a median Dice similarity coefficient of 0.86. Recently, Bulman et al. [27] compared prostate volumes obtained by automated computer-generated multifeature active shape models based on MRI with pathology specimen volumes in 91 patients and reported that multifeature active shape models had the highest slope (0.888) with a concordance correlation coefficient of 0.867 with respect to prostatectomy specimen volume. However, this system requires the operator to identify the center of the prostate on a single midgland axial T2-weighted section; thus, it is not completely automated. The method used in our study differs from prior approaches in that it includes a 3D approach and uses normalized gradient fields cross-correlation and a graph-based search. It
yielded reasonable and encouraging correlation with both true prostate volume of the specimens and with manual segmentation volumes in a fully automatic and highly time-efficient (≈ 10 seconds) manner without the need for cursor placement by the user. Thus, this method avoids subjective differences among different viewers and yields a highly reproducible result. The best prostate volume estimates obtained with manual segmentation were quite close to those obtained with the fully automated system; for example, the Pearson correlation coefficient between manual and automated axial prostate volume estimates and true prostate volume (from prostatectomy specimen) were 0.90, whereas for manual and automated TPM 0.5 prostate volume were 0.91 and 0.90, respectively. The partial mean Dice similarity coefficient (which excludes slices that were not annotated) for triplane automated segmentations ranged between 0.90 and 0.92, whereas the full mean Dice similarity coefficient (which includes all slices) ranged between 0.83 and 0.89, with 0.89 representing the axial full mean Dice similarity coefficient. Thus, this automated prostate segmentation tool can provide a convenient way to estimate prostate volume and to segment the prostate, which can potentially be used in clinical management of prostate cancer patients and in research protocols. Our study has several limitations. First, we used fresh prostatectomy specimen weight as the ground truth to validate our results and the mean specimen volume was higher than either
AJR:201, November 2013 W723
Turkbey et al. TABLE 4: Partial and Full Dice Similarity Coefficient Values Between Manually and Fully Automated Segmented Single-Plane and Thresholded Probability Map (TPM) Shapes Dice Similarity Coefficient Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Shapes
Mean
SD
Standard Error
Partial Dice similarity coefficient Axial
0.92
0.03
0.003
Sagittal
0.90
0.03
0.003
Coronal
0.91
0.03
0.003
Axial
0.89
0.04
0.004
Sagittal
0.87
0.04
0.004
Coronal
0.83
0.06
0.006
TPM 0.7
0.89
0.04
0.004
Full Dice similarity coefficient
TPM 0.5
0.85
0.05
0.005
TPM 0.3
0.90
0.03
0.003
the manual or automated volume estimates because the seminal vesicles were not amputated from the specimen before weighing. Moreover, variable amounts of extraprostatic tissue may be included in some specimens when the surgeon decided to perform a wider resection around the prostate. There was no reliable way to correct for this discrepancy; however, we believe that the seminal vesicles and minimal periprostatic fat contribute modestly to the weight of the prostate. It is also likely, however, that the ex vivo specimen is somewhat smaller because of the loss of blood from the gland. This problem in comparisons with ex vivo tissue is unavoidable. An additional issue is that the manual segmentations were performed by a single experienced operator. However, this reader’s results correlated well with the histopathology specimen and we believe that those results are reliable. Third, the MR images were obtained with an endorectal coil, which compresses the gland posteriorly, potentially affecting the volume estimation of the prostate compared with specimens [28]. We believe that this effect is small and the value of the endorectal coil is that it provides images with superior resolution for delineating the prostate boundaries, which is critical for accurate volume determinations. Anecdotally, this automated segmentation system performs well on images obtained without an endorectal coil. Thus, on balance, this limitation probably contributes minimally to disparities between pathologic specimens and MRI-derived volumes. In conclusion, 3D segmentations of the prostate and prostate volumes derived from using an automated segmentation tool based on nor-
W724
malized gradient fields, cross-correlation, and graph-search refinement can yield accurate prostate volume determinations in a remarkably time-efficient manner. This time-consuming task can potentially be used routinely because it requires essentially no user input and only 10 seconds to complete. Applications of accurate prostate segmentation go beyond simple volume determinations, PSA density measurements, and follow-up of benign prostate hyperplasia and extend to multimodal image fusion and multi–time point image fusion with implications for automated detection, biopsy, and imaging-guided therapy. References 1. Ramon J, Boccon-Gibod L, Billebaud T, et al. Prostate-specific antigen density: a means to enhance detection of prostate cancer. Eur Urol 1994; 25:288–294 2. Blackwell KL, Bostwick DG, Myers RP, Zincke H, Oesterling JE. Combining prostate specific antigen with cancer and gland volume to predict more reliably pathological stage: the influence of prostate specific antigen cancer density. J Urol 1994; 151:1565–1570 3. Stephan C, Cammann H, Meyer HA, Lein M, Jung K. PSA and new biomarkers within multivariate models to improve early detection of prostate cancer. Cancer Lett 2007; 249:18–29 4. Corcoran NM, Casey RG, Hong MK, et al. The ability of prostate-specific antigen (PSA) density to predict an upgrade in Gleason score between initial prostate biopsy and prostatectomy diminishes with increasing tumour grade due to reduced PSA secretion per unit tumour volume. BJU Int 2012; 110:36–42
5. Choi SY, Chang IH, Kim YS, Kim TH, Kim W, Myung SC. Prostate specific antigen velocity per prostate volume: a novel tool for prostate biopsy prediction. Urology 2011; 78:874–879 6. Ko JS, Landis P, Carter HB, Partin AW. Effect of intra-observer variation in prostate volume measurement on prostate-specific antigen density calculations among prostate cancer active surveillance participants. BJU Int 2011; 108:1739–1742 7. Sfoungaristos S, Perimenis P. PSA density versus risk stratification for lymphadenectomy-making decision in patients with prostate cancer undergoing radical prostatectomy. Int Urol Nephrol 2011; 43:1073–1079 8. Eskicorapci SY, Guliyev F, Akdogan B, Dogan HS, Ergen A, Ozen H. Individualization of the biopsy protocol according to the prostate gland volume for prostate cancer detection. J Urol 2005; 173:1536–1540 9. Al-Qaisieh B, Ash D, Bottomley DM, Carey BM. Impact of prostate volume evaluation by different observers on CT-based post-implant dosimetry. Radiother Oncol 2002; 62:267–273 10. Vikal S, Haker S, Tempany C, Fichtinger G. Prostate contouring in MRI guided biopsy. Proc SPIE 2009; 7259:72594A 11. Giubilei G, Ponchietti R, Biscioni S, et al. Accuracy of prostate volume measurements using transrectal multiplanar three-dimensional sonography. Int J Urol 2005; 12:936–938 12. Matthews GJ, Motta J, Fracehia JA. The accuracy of transrectal ultrasound prostate volume estimation: clinical correlations. J Clin Ultrasound 1996; 24:501–505 13. Rahmouni A, Yang A, Tempany CM, et al. Accuracy of in-vivo assessment of prostatic volume by MRI and transrectal ultrasonography. J Comput Assist Tomogr 1992; 16:935–940 14. Sosna J, Rofsky NM, Gaston SM, DeWolf WC, Lenkinski RE. Determinations of prostate volume at 3-Tesla using an external phased array coil: comparison to pathologic specimens. Acad Radiol 2003; 10:846–853 15. Smith WL, Lewis C, Bauman G, et al. Prostate volume contouring: a 3D analysis of segmentation using 3DTRUS, CT, and MR. Int J Radiat Oncol Biol Phys 2007; 67:1238–1247 16. Lee JS, Chung BH. Transrectal ultrasound versus magnetic resonance imaging in the estimation of prostate volume as compared with radical prostatectomy specimens. Urol Int 2007; 78:323–327 17. Jeong CW, Park HK, Hong SK, Byun SS, Lee HJ, Lee SE. Comparison of prostate volume measured by transrectal ultrasonography and MRI with the actual prostate volume measured after radical prostatectomy. Urol Int 2008; 81:179–185 18. Fotin SV, Yin Y, Periaswamy S, et al. Normalized gradient fields cross-correlation for automated de-
AJR:201, November 2013
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI tection of prostate in magnetic resonance images. Proc SPIE 2012; 8314:83140V: 19. Yin Y, Fotin SV, Periaswamy S, et al. Fully automated prostate segmentation in 3D MR based on normalized gradient field cross-correlation initialization and LOGISMOS refinement. Proc SPIE 2012; 8314:831406 20. Biancardi AM, Jirapatnakul AC, Reeves AP. A comparison of ground truth estimation methods. Int J Comput Assist Radiol Surg 2010; 5:295–305 21. Rodriguez E Jr, Skarecky D, Narula N, Ahlering TE. Prostate volume estimation using the ellipsoid formula consistently underestimates actual gland size. J Urol 2008; 179:501–503 22. Dice LR. Measures of the amount of ecologic asso-
ciation between species. Ecology 1945; 26:297–302 23. Jia G, Baudendistel KT, von Tengg-Kobligk H, et al. Assessing prostate volume by magnetic resonance imaging: a comparison of different measurement approaches for organ volume analysis. Invest Radiol 2005; 40:243–248 24. Pasquier D, Lacornerie T, Vermandel M, Rousseau J, Lartigau E, Betrouni N. Automatic segmentation of pelvic structures from magnetic resonance images for prostate cancer radiotherapy. Int J Radiat Oncol Biol Phys 2007; 68:592–600 25. Klein S, van der Heide UA, Lips IM, van Vulpen M, Staring M, Pluim JP. Automatic segmentation of the prostate in 3D MR images by atlas matching using localized mutual information. Med Phys
2008; 35:1407–1417 26. Martin S, Troccaz J, Daanenc V. Automated segmentation of the prostate in 3D MR images using a probabilistic atlas and a spatially constrained deformable model. Med Phys 2010; 37:1579–1590 27. Bulman JC, Toth R, Patel AD, et al. Automated computer-derived prostate volumes from MR imaging data: comparison with radiologist-derived MR imaging and pathologic specimen volumes. Radiology 2012; 262:144–151 28. Heijmink SW, Scheenen TW, van Lin EN, et al. Changes in prostate shape and volume and their implications for radiotherapy after introduction of endorectal balloon as determined by MRI at 3T. Int J Radiat Oncol Biol Phys 2009; 73:1446–1453
APPENDIX 1: Mass and Volume Estimation Errors In another experiment we estimated how accurately we can predict true prostate mass from the prostate volume obtained using MR images. The quality of such an estimate depends on several factors such as the accuracy of the segmentation, geometric distortion caused by MRI, and minor variations in the density of the prostate tissue. The relative error in the prostate mass estimate is the upper bound of the error in physical volume estimate because the density variation does not have an impact on the volume. To measure the mass estimation error, we split available studies into training and test subsets. From the training subset, we constructed two models: The objective of the first model was to predict the true mass m using the volume Vm from manually traced axial contours and then using the volume from the automated segmentation (Va) obtained with thresholded probability map (TPM) 0.5 method. For modeling we used linear regression in the following form: m = α + V + β, where α is a multiplicative component that roughly represents prostate tissue density and β is an additive component to account for variation in measured mass due to the seminal vesicles and the excess of the extracted tissue.
Localization
Linear regression parameters were estimated using least squares over the training set and the following coefficients were established for MRI-derived segmentations: α = 0.820 g/cm3 and β = 18.680 g for manual segmentation and α = 0.804 g/cm3 and β = 17.830 g for automated segmentation. These linear regression models were then applied to each of the studies in the test subset and a mass estimation error was calculated from measured prostate mass m and estimated mass me. The absolute error (Δm = me – m) and relative error (δm = Δm / m) were calculated for each of the studies. For manually obtained segmentation, the absolute and relative errors were in the range of from –25.92 to 11.39 g and from –30.03% to 30.13%, respectively. For 95% of the studies, the estimated mass fell within 28.25% of the measured mass. The root mean squared error for manual segmentation was 13.90%. For automatically obtained segmentation, the absolute and relative errors were in the range of from –20.45 to 9.76 g and from –32.26% to 31.38%, respectively. Here for 95% of the studies the estimated mass fell within 25.73% of the measured mass. The root mean squared error for automatic segmentation was 13.10%. Thus, MR images can also be used to effectively estimate the prostate mass.
Refinement
Fig. 1—Automated prostate segmentation consists of two steps: localization and refinement. Red shows slice of 3D prostate bounding box as identified by localization step, and green shows evolution of prostate surface during refinement step.
AJR:201, November 2013 W725
Turkbey et al.
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Fig. 2—Prostate segmentation visualization modes. A, Three-dimensional surface rendering shows prostate (green). B, Two-dimensional cross sections were overlaid on original MR image slice to produce this image. Prostate is outlined in green.
A
B Fig. S3—Screen shot from video of fully automated segmentation tool (Medical Image Processing, Analysis and Visualization [MIPAV], Center for Information Technology, National Institutes of Health) in sample case. (To view this video, click Supplemental at the top of this article and then click the figure number on the Supplemental page.)
A
W726
B
Fig. 4—Thresholded probability map. A, Sketch shows original three segmentations. B, Sketch shows probability map thresholded at level of 0.5.
AJR:201, November 2013
A
B
C
Fig. 5—Partial Dice similarity coefficient. A, Axial MR image with manually drawn contour showing prostate (green). B, Sagittal MR image shows cross sections (green lines) of multiple manually drawn axial contours. C, Sagittal MR image shows portions of prostate removed (dashed lines) for calculation of partial Dice similarity coefficient. Green shows outline of prostate.
80 70
Manual segmentation Automated segmentation True prostate volume
80
60
70
50
Mean Prostate Volume (cm3)
Mean Prostate Volume (cm3)
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI
40 30 20 10 0
Axial
Coronal
Sagittal
Manual segmentation Automated segmentation True prostate volume
60 50 40 30 20 10
Mean of Single-Plane Values
0
TPM 0.7
TPM 0.5
TPM 0.3
A
B
Fig. 6—Bar graphs show mean prostate volume estimates and mean true prostate volume. Vertical lines show mean ± 1 SD. A, Bar graph shows prostate volume estimates based on manual segmentation and automated segmentation and true prostate volumes extracted from prostatectomy specimens in three imaging planes and mean values overall. B, Bar graph shows prostate volume estimates based on thresholded probability map (TPM) approaches at 0.7, 0.5, and 0.3 and true prostate volumes extracted from prostatectomy specimens.
AJR:201, November 2013 W727
D
G
180 160 140 120 100 80 60 40 20 0
150 50 100 PV in Axial Plane Using Automated Segmentation (cm3)
150 50 100 PV in Axial Plane Using Manual 3 Segmentation (cm )
20
B 160 140 120 100 80 60 40 20 0
200
0
200
H
180 160 140 120 100 80 60 40 20 0
40 60 80 100 120 140 160 Mean PV Using Manual Segmentation (cm3)
50 100 150 Mean PV Using Automated Segmentation (cm3)
20
40 60 80 100 120 140 160 Mean PV Using Manual Segmentation (cm3)
y = 0.8948x + 17.857 R 2 = 0.8211
C 160 140 120 100 80 60 40 20 0
200
y = 1.0276x + 2.5995 R 2 = 0.9475
0
160 140 120 100 80 60 40 20 0 0
y = 0.8459x + 16.963 R 2 = 0.81
E
y = 0.9744x + 1.2638 R 2 = 0.9468
0
True PV (cm3) 0
y = 0.852x + 18.116 R 2 = 0.8163
0
y = 0.8961x + 18.16 R 2 = 0.8172
True PV (cm3)
160 140 120 100 80 60 40 20 0
200
True PV (cm3)
True PV (cm3)
A
50 100 150 PV in Axial Plane Using Manual 3 Segmentation (cm )
160 140 120 100 80 60 40 20 0
F
I
180 160 140 120 100 80 60 40 20 0
20 40 60 80 100 120 140 160 PV Using Manual Segmentation With TPM 0.5 (cm3)
y = 0.8464x + 17.724 R 2 = 0.8044
0
PV Using Automated Segmentation With TPM 0.5 (cm3)
True PV (cm3)
y = 0.8541x + 18.291 R 2 = 0.8183
Mean PV Using Automated Segmentation (cm3)
True PV (cm3)
160 140 120 100 80 60 40 20 0 0
PV in Axial Plane Using Automated Segmentation (cm3)
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Turkbey et al.
100 150 50 PV Using Automated Segmentation With TPM 0.5 (cm3)
200
y = 1.0231x + 1.439 R 2 = 0.9548
0
20 40 60 80 100 120 140 160 PV Using Manual Segmentation With TPM 0.5 (cm3)
Fig. 7— Linear regression plots. A–I, Linear regression plots show prostate volume (PV) estimates derived from manual and automated segmentations and true prostate volumes extracted from prostatectomy specimens. TPM = thresholded probability map.
W728
AJR:201, November 2013
Downloaded from www.ajronline.org by 203.142.40.243 on 11/04/15 from IP address 203.142.40.243. Copyright ARRS. For personal use only; all rights reserved
Automated Prostate Segmentation on MRI
A
B
C
D
E
F
Fig. 8—Sample segmentations. A–F, Axial (A and B), sagittal (C and D), and coronal (E and F) MR images show manual tracings (green) and automatically generated segmentations (red) of prostate. A, C, and E are images of 62-year-old man and B, D, and F are images of 56-year-old man.
F O R YO U R I N F O R M AT I O N
The data supplement accompanying this web exclusive article can be viewed by clicking “Supplemental” at the top of the article.
AJR:201, November 2013 W729
