ORIGINAL RESEARCH
n GENITOURINARY IMAGING
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 www.rsna.org/rsnarights.
Automated Computer-derived
Prostate Volumes from MR
Imaging Data: Comparison with
Radiologist-derived MR Imaging and
Pathologic Specimen Volumes1
Julie C. Bulman, BA
Robert Toth, MS
Amish D. Patel, MD
B. Nicolas Bloch, MD
Colm J. McMahon, MB
Long Ngo, PhD
Anant Madabhushi, PhD
Neil M. Rofsky, MD
Purpose:
To compare prostate gland volume (PV) estimation of automated computer-generated multifeature active shape
models (MFAs) performed with 3-T magnetic resonance
(MR) imaging with that of other methods of PV assessment,
with pathologic specimens as the reference standard.
Materials and
Methods:
All subjects provided written informed consent for this
HIPAA-compliant and institutional review board–approved
study. Freshly weighed prostatectomy specimens from
91 patients (mean age, 59 years; range, 42–84 years)
served as the reference standard. PVs were manually calculated by two independent readers from MR images by
using the standard ellipsoid formula. Planimetry PV was
calculated from gland areas generated by two independent
investigators by using manually drawn regions of interest.
Computer-automated assessment of PV with an MFA was
determined by the aggregate computer-calculated prostate area over the range of axial T2-weighted prostate MR
images. Linear regression, linear mixed-effects models,
concordance correlation coefficients, and Bland-Altman
limits of agreement were used to compare volume estimation methods.
Results:
MFA-derived PVs had the best correlation with pathologic
specimen PVs (slope, 0.888). Planimetry derived volumes
produced slopes of 0.864 and 0.804 for two independent
readers when compared with specimen PVs. Ellipsoid
formula–derived PVs had slopes closest to one when compared with planimetry PVs. Manual MR imaging and MFA
PV estimates had high concordance correlation coefficients with pathologic specimens.
Conclusion:
MFAs with axial T2-weighted MR imaging provided an automated and efficient tool with which to assess PV. Both
MFAs and MR imaging planimetry require adjustments for
optimized PV accuracy when compared with prostatectomy
specimens.
1
From the Georgetown University School of Medicine,
Washington, DC (J.C.B.); Department of Radiology, Beth
Israel Deaconess Medical Center, Boston, Mass (A.D.P.,
C.J.M., L.N.); Laboratory for Computational Imaging and
Bioinformatics, Rutgers University, Piscataway, NJ (R.T.,
A.M.); Department of Radiology, Boston University School
of Medicine, Boston, Mass (B.N.B.); and Department of
Radiology, University of Texas Southwestern Medical
Center, 5323 Harry Hines Blvd, Dallas, TX 75390 (N.M.R.).
Received February 15, 2011; revision requested April 11;
revision received July 25; accepted August 10; final version
accepted August 22. Supported by grants from the Wallace
H. Coulter Foundation, the New Jersey Commission on
Cancer Research, the Cancer Institute of New Jersey, and
Bioimagene. Address correspondence to N.M.R. (e-mail:
neil.rofsky@utsouthwestern.edu ).
q
q
RSNA, 2012
RSNA, 2012
144
radiology.rsna.org
n
Radiology: Volume 262: Number 1—January 2012
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
A
determination of prostate gland
volume (PV) facilitates an assessment of prostate disorders and,
for prostate cancer, in conjunction with
other parameters, can help predict the
pathologic stage of disease, offer insights into the prognosis, and help
predict treatment response (1,2). It
can facilitate therapeutic planning for
brachytherapy, cryotherapy, and minimally invasive benign prostatic hyperplasia therapy (3–7). Prostate-specific
antigen levels have been modified to derive the prostate-specific antigen density
by incorporating PV calculations to help
guide clinical decisions (8,9). The clinical value of prostate-specific antigen
density, however, is dependent on the
quality of the PV estimate. The accuracy
and variability of PV determinations pose
limitations to its usefulness in clinical
practice (5,10,11).
There are a variety of methods for PV
estimation in terms of imaging modality and the strategy used to determine
the volume (8,9,12–14). It is generally
accepted that step-section planimetry
yields better results compared with
formula calculations and that magnetic
resonance (MR) imaging can yield reasonably accurate results, often better
than those of transrectal ultrasonography (US) and computed tomography
(CT) (4).
Planimetry with MR imaging has
been shown to be the most accurate
method with which to determine PV
to date, but it is often too time consuming for efficient clinical practice; thus,
alternative formula-derived methods are
commonly used (13,15–17). While the
ellipsoid formula is used as a practical
clinical protocol for time-efficient PV
estimation, it can yield substantial inaccuracies with two-dimensional transrectal US (11,18). Other formulas used
to calculate prostate gland size are
available, but to our knowledge there is
no universally accepted approach that
yields accurate results for the diverse
sizes and shapes of prostate glands encountered in clinical practice or that
helps to account for the influence of
operator experience on volume determination (19).
In our study, we explored the use
of fully automated multifeature active
shape models (MFAs) for PV measurement with the goal of identifying a reliable, reproducible, and efficient tool
for clinical practice. This MFA is based
on a traditional segmentation system
that uses active shape models. Prostate
active shape models learn the shape
of the prostate given a set of training
studies and learn to define the prostate border by using the adjacent signal
intensities. In addition, our MFA uses
multiple statistical texture features to
drive the appearance model, instead
of simply using signal intensities, in an
attempt to yield more accurate volume
estimations. Active shape models have
also been used to detect prostate cancer by using MR spectroscopy and T2weighted ex vivo prostate MR imaging
(20–23). The purpose of our study was,
therefore, to compare MFA volume estimations to ellipsoid formula-derived
and planimetry derived volumes, with
pathologic specimens as the reference
standard.
Advances in Knowledge
n Multifeature active shape models
(MFAs) offer a computerized and
automated alternative with which
to determine in vivo prostate
gland volumes (PVs).
n MFA distribution of PV estimates
is similar to that of planimetry,
has a high concordance correlation coefficient with planimetry,
and thus, may be an acceptable
method with which to determine PV.
Radiology: Volume 262: Number 1—January 2012
n
Materials and Methods
This retrospective study was an arm of
a prospective study that was approved
by the institutional review board of Beth
Implication for Patient Care
n Consistently and accurately
determined PVs obtained by
using an MFA may facilitate appropriate clinical decisions that
are based on PV determination.
radiology.rsna.org
Bulman et al
Israel Deaconess Medical Center and
was compliant with the Health Insurance Portability and Accountability Act.
All subjects provided written informed
consent.
From August 2007 to May 2009,
96 consecutive subjects with wholemounted specimens from radical prostatectomy who had undergone preoperative pelvic MR imaging were initially
included in our study. Two subjects were
excluded for lack of access to MR images. One subject was excluded because
MR imaging was performed with a 1.5T magnet. One subject was excluded
for lack of endorectal coil use. One subject was excluded because pathologic
specimen weight was unavailable. A total
of 91 patients (mean age, 59 years; age
range, 42–84 years) were included in
this data set (Fig 1).
Reference Standard
Ninety-one prostatectomy specimens
(prostate gland with attached seminal
vesicles), which were removed during
radical prostatectomy, were weighed
by a pathologist when fresh. As demonstrated by Rodriguez et al (24), there
Published online
10.1148/radiol.11110266 Content codes:
Radiology 2012; 262:144–151
Abbreviations:
CI = confidence interval
MFA = multifeature active shape model
PV = prostate gland volume
ROI = region of interest
Author contributions:
Guarantors of integrity of entire study, J.C.B., R.T., A.M.,
N.M.R.; study concepts/study design or data acquisition
or data analysis/interpretation, all authors; manuscript
drafting or manuscript revision for important intellectual
content, all authors; manuscript final version approval,
all authors; literature research, J.C.B., R.T., A.D.P., B.N.B.,
A.M.; clinical studies, J.C.B., B.N.B., C.J.M., A.M., N.M.R.;
statistical analysis, J.C.B., L.N., A.M.; and manuscript editing, J.C.B., R.T., A.D.P., C.J.M., L.N., A.M., N.M.R.
Funding:
This research was supported by the National Institute of
Health (grant 5R01CA116465-03) and the National Cancer
Institute (grants R01CA136535-01, R01CA14077201,
R21CA12718601, and R03CA143991-01).
Potential conflicts of interest are listed at the end
of this article.
145
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
Bulman et al
Figure 1
were then drawn section by section
up to the base and down to the apex.
Upon completing the ROI tracing, the
software displays the area for each section in square centimeters. The volume
was estimated by multiplying the sum of
these areas by the section thickness in
centimeters.
Figure 1: Flowchart of subjects included in the study (n = 91) and reasons
for exclusion. ERC = endorectal coil.
is a 0.997 correlation between prostate
weight and displaced water volume in
milliliters; Varma and Morgan (25) found
a similar agreement. Therefore, the specimen weight is used as the true volume
of the prostate gland. Rodriguez et al
(24) also showed 3.8 g to be the average weight of seminal vesicles, and this
number was subtracted from specimen weight to compensate for the attached seminal vesicles. This calculated
weight multiplied by 1.05 g/mL (specific
gravity of prostatic tissue) served as the
reference standard for our study (26).
MR Imaging Technique
Ninety-one prostates were analyzed with
MR images acquired with a 3-T wholebody imager (GE Healthcare Technologies, Waukesha, Wis) and an endorectal coil (Medrad, Pittsburgh, Pa) inflated
with 60 mL of a 100% weight per volume
barium sulfate suspension for improved
spatial resolution (27–29). T2-weighted
MR images were acquired in the axial,
coronal, and sagittal planes (repetition time msec/echo time msec, 3300–
6250/155–165; echo train length, 20–21;
four signals acquired; field of view,
16 3 16 cm [coronal and sagittal] or
14 3 14 cm [axial]; section thickness,
2.2–3.0 mm with no gap). The section thickness varied according to the
size of the prostate gland to maintain
a consistent imaging time and in-plane
resolution.
PV Estimation with the Ellipsoid Formula
Total PVs were calculated from T2weighted axial MR images in a prospec146
tive real-time reading fashion as part
of the routine clinical interpretation by
members of the clinical radiology team
(seven fellowship-trained radiologists
[including C.J.M.], with 1–20 years of
experience reading prostate MR images,
who are hereafter collectively referred
to as reader 1) and retrospectively by a
radiologist (B.N.B., with 9 years experience interpreting prostate MR imaging,
hereafter referred to as reader 2) who
was not involved in any of the prospective clinical interpretations of reader 1.
Readers calculated the PVs by using the
standard ellipsoid formula: DAP · DML ·
DTV · p /6, where DAP is the maximum
anteroposterior dimension measured
on sagittal images, DML is the maximum
craniocaudal dimension measured on
sagittal images, and DTV is the maximum transverse dimension measured on
axial images.
PV Estimation with Planimetry
To establish a surrogate in vivo ground
truth, planimetry was performed by two
independent readers (reader 3: A.D.P.,
with less than 1 year of experience;
reader 4: C.J.M., with more than 6
years of experience) who did not have
prior exposure to the cases. T2-weighted
axial MR images of the prostate were
analyzed by using a workstation (Advantage 4.0; GE Medical Systems). The
prostatic capsule was manually traced
by drawing a region of interest (ROI)
with the cursor on each two-dimensional
MR image of the series, which took
5–10 minutes per case. ROIs were initially drawn around the midgland and
PV Estimation with MFA
Computer automated assessment of PV
was determined by using noncommercial software that analyzed the aggregate computer-calculated prostate area
over the range of the 91 test set axial
T2-weighted MR imaging cases. This
volume calculation is a single-click automated process that requires a maximum of 200 seconds per prostate gland.
An extension to traditional active shape
models (30) was implemented, which
used statistical texture features (mean,
variance, gradient magnitude) instead
of the signal intensities of the image to
determine the location of the prostate,
which was called an MFA. The method
was similar to that used by Toth et al
(31), except it was performed in three
dimensions. A single click in the center
of the prostate on a single midgland T2weighted axial section was used to initialize the MFA for propagation through
the entire data set, and the locations
that were determined to lie on the surface of the prostate were detected from
the texture features. The number of
voxels enclosed by this surface was
thus used to estimate the volume of the
prostate (Fig 2). Thirty-one cases obtained with 3-T T2-weighted endorectal
coil MR imaging (mean PV, 52.1 mL;
range, 25–112 mL), which were distinct
from the test set, were used to train
the MFA once in an offline setting. The
MFA was trained by placing a series of
equally spaced landmarks on the surface of the organ. Principal component
analysis was then used to capture shape
variations, such that the segmentation
can be constrained to only valid prostate shapes (prostate shapes similar to
the training set). To train the appearance of the prostate border, a Gaussian distribution of the texture features
surrounding each landmark was created. Hence, each landmark has a
radiology.rsna.org
n
Radiology: Volume 262: Number 1—January 2012
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
distinctive appearance model, allowing
for variations of appearance at different locations of the prostate surface.
To segment the prostate in a new image, the same set of textural features
is extracted, and the locations with
the highest probability of lying on the
surface are detected. However, the locations for which the textural features
yield the highest probability might not
all be accurate, or the resulting surface
might be jagged. The trained statistical
shape model is then used to constrain
and smooth the segmentation to only
valid prostate shapes; thus, the prostate is segmented. This method was
previously used for successful volume
estimations of the prostate by Toth et al
(32), where the MFA-derived volumes
were compared with manually estimated volumes obtained by planimetric
estimation of the prostate capsule derived from manual segmentations by an
expert radiologist on T2-weighted MR
images.
Statistical Analysis
The Wilcoxon signed rank test was used
to compare the mean PV between any
two methods. There were 15 pairwise
comparisons of interest (Table 1); therefore, type I error was adjusted by using
the Bonferroni multiple comparison adjustment (a level of .05 divided by 15
comparisons, yielding an adjusted a
of .003). The P values were compared
with multiple comparison–adjusted type
I error. To assess concordance between
two methods, linear regression was
used to obtain the individual slope and
its 95% CI. To test for the difference
between any two of these slopes (there
were 10 comparisons of interest of five
slopes [Table 1]), linear mixed-effects
models (33) with linear contrasts with
compound symmetry structure for the
variance-covariance were used. In addition, the concordance correlation coefficient and its 95% CI were computed. The percentage of measurements
whose between-methods differences were
within the limits of agreement from the
Bland-Altman plot was also reported.
All analyses were performed with SAS
software (version 8; SAS Institute,
Cary, NC).
Radiology: Volume 262: Number 1—January 2012
n
Bulman et al
Figure 2
Figure 2: Left: Axial T2-weighted MR image. Right: Derived prostate volume. Blue = MFA-derived volume,
green = planimetry-derived volume.
Results
Distribution of PVs
Table 2 shows the mean 6 standard deviation, median, and range for PV with
each of the six methods. The distributions
show asymmetry, with the medians consistently smaller than the means, indicating slight skewness of the right tail. The
mean and median PVs with the four MR
imaging methods and MFA are consistently smaller than those with the pathologic reference standard. This indicates a
potential underestimation of the PV with
the MR imaging and MFA methods. The
variability in the distribution of these five
methods is similar to that of the pathologic reference standard. The P values
from Table 1 for comparison of these five
methods with the pathologic reference
standard indicate that the underestimation is indeed significant (15.8%, P =
.0001). The first column in Table 1 also
shows the mean amount and variability
of underestimation.
PV Estimates with the Ellipsoid Formula
The ellipsoid formula as performed
by reader 1 was the least correlated
with the pathologic reference standard
(slope, 0.805; 95% CI: 0.707, 0.903).
The ellipsoid formula as performed by
reader 2 had a slope of 0.864 (95%
CI: 0.786, 0.942) when compared with
pathologic reference standard. The ellipsoid formula as performed by reader
1 had slopes of 0.910 and 0.987 when
radiology.rsna.org
compared with planimetry as performed
by readers 3 and 4, respectively. When
compared with each other, calculations
by readers 1 and 2 had a slope of 0.897
(95% CI: 0.810, 0.983). The underestimation of the PV as compared with
the pathologic reference standard was
similar for readers 1 and 2 (29.56 vs
27.58). Reader 2 also had a higher concordance correlation coefficient (0.857
vs 0.779) and percentage of measurements within the limits of agreement
(95.6% vs 92.3%).
PV Estimates with Planimetry
This analysis was performed to assess
whether planimetry was an appropriate surrogate ground truth for PV obtained with the pathologic reference
standard. The slope of the regression
model in comparison to the pathologic
specimens was 0.864 for reader 3 and
0.804 for reader 4. When the two planimetry data sets were compared, the
slope of the line was 1.074 (95% CI:
1.046, 1.101). Reader 4 had a higher
concordance correlation coefficient (0.897
vs 0.843) and a similar percentage of
measurements within the Bland-Altman
limits of agreement (95.6% vs 94.5%)
compared with reader 3. The mean
amount of underestimation of the pathologic PV was smaller in reader 4 than
in reader 3 (25.57 vs 29.26).
PV Estimates with MFA
Pathologically determined PV was underestimated by a mean 27.76 6 7.69
147
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
Bulman et al
Table 1
Linear Regression Data for Clinical and MFA PV Measurement Methods
Comparison
Ellipsoid formula vs pathologic standard
Reader 1
Reader 2
Planimetry vs pathologic standard
Reader 3
Reader 4
MFA vs pathologic standard
MFA vs ellipsoid formula
Reader 1
Reader 2
Ellipsoid formula vs planimetry
Reader 1 vs reader 3
Reader 1 vs reader 4
Reader 2 vs reader 3
Reader 2 vs reader 4
Planimetry vs MFA
Reader 3
Reader 4
Planimetry: reader 3 vs reader 4
Ellipsoid formula: reader 1 vs reader 2
P Value†
Linear Regression
Equation
95% CI of
the Slope
Concordance Correlation
Coefficient‡
Measurements
within Limits of
Agreement (%)§
29.56 6 10.88
27.58 6 8.49
.0001
.0001
y = 0.805x 1 17.51||
y = 0.864x 1 13.39||
0.707, 0.903
0.786, 0.942
0.779 (0.707, 0.852)
0.857 (0.807, 0.907)
92.3
95.6
29.26 6 7.92
25.57 6 8.34
27.76 6 7.69
.0001
.0001
.0001
y = 0.864x 1 14.85||
y = 0.804x 1 14.32||
y = 0.888x 1 12.55||
0.793, 0.935
0.741, 0.867
0.800, 0.976
0.843 (0.792, 0.895)
0.897 (0.858, 0.935)
0.867 (0.821, 0.913)
94.5
95.6
93.4
1.79 6 12.36
20.18 6 10.29
.225
.700
y = 0.849x 1 4.65
y = 0.891x 1 4.82
0.730, 0.967
0.791, 0.990
0.828 (0.763, 0.893)
0.882 (0.836, 0.928)
95.6
93.4
20.30 6 9.22
24.00 6 9.61
1.68 6 5.34
22.01 6 5.35
.756
.105
.381
.0001
y = 0.910x 1 3.975
y = 0.987x 1 4.502
y = 0.983x 2 0.935
y = 1.070x 2 0.990
0.822, 0.998
0.893, 1.081
0.930, 1.036
1.019, 1.121
0.908 (0.872, 0.944)
0.894 (0.853, 0.935)
0.966 (0.952, 0.980)
0.967 (0.954, 0.981)
93.4
96.7
93.4
92.3
21.50 6 9.38
2.19 6 9.70
23.69 6 3.25
21.98 6 9.17
.265
.027
.162
.222
y = 0.879x 1 6.475
y = 0.818x 1 5.955
y = 1.074x 1 0.661
y = 0.897x 1 6.190
0.791, 0.967
0.739, 0.896
1.046, 1.101
0.810, 0.983
0.900 (0.861, 0.939)
0.900 (0.861, 0.939)
0.976 (0.967, 0.985)
0.904 (0.866, 0.942)
93.4
93.4
95.6
94.5
Mean
Difference*
Note.—CI = confidence interval.
* Data are means 6 standard deviations.
†
Wilcoxon signed rank test for paired data used to compare differences in mean PV. Bonferroni-adjusted type I error set at .003 (.05 divided by 15 comparisons).
‡
Data in parentheses are 95% CIs.
§
Limits of agreement were the upper and lower bounds from the Bland-Altman plot (bias 6 1.96*SD of differences).
||
Ten pairwise slopes were compared by using a linear mixed-effects model, with type I error set at .003. Significant differences were found between the ellipsoid formula as performed by reader 1
and planimetry as performed by reader 4 (P = .001), the ellipsoid formula as performed by reader 2 and planimetry as performed by reader 4 (P = .0001), and planimetry between readers 3 and 4 (P
= .0001).
with MFA. Volume estimates with MFA
yielded the slope closest to one when
compared with the pathologic reference
standard (slope, 0.888; [95% CI: 0.800,
0.976]; concordance correlation coefficient, 0.867) (Fig 3). MFA PVs were
compared with planimetry PVs; the results were not significant for readers 3 or
4 (P = .265 and .027, respectively) with
the adjusted type I error level of .004.
The concordance correlation coefficient
between MFA and the pathologic reference standard was 0.867, with 93.4%
of the measured differences within the
Bland-Altman limits of agreement.
Discussion
The computer segmentation method used
in our study, MFA, brings automation
and accuracy to volume determinations.
148
Table 2
PVs as Estimated with Clinical and MFA Methods
Method
Ellipsoid formula
Reader 1
Reader 2
Planimetry
Reader 3
Reader 4
MFA
Pathologic reference standard
Mean (mL)*
Median (mL)
Range (mL)
40.81 6 21.38
42.80 6 21.14
35.40
38.58
14.57–128.11
13.23–155.77
41.11 6 21.44
44.81 6 23.20
42.62 6 20.89
50.38 6 19.94
35.22
39.22
37.75
45.36
12.72–139.13
14.70–151.94
12.54–133.41
23.31–138.81
* Data are means 6 standard deviations.
Manual placement of ROIs and cursors
is inherently user dependent and, thus,
introduces variability into manually
generated volume determinations. The
expected improved consistency from
automated volume determinations offers
the potential to improve clinical usefulness of PVs.
MR imaging yields strong soft-tissue
contrast and a more accurate assessment
radiology.rsna.org
n
Radiology: Volume 262: Number 1—January 2012
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
Figure 3
Figure 3: Graph shows MFA PV estimates in comparison to pathologic reference standard
(Path) PVs. cc = mL, R2 = coefficient of determination.
of PV than does transrectal US or CT
(12,13,34). MR imaging reduces intraobserver inconsistency by a factor of
3.5, at a minimum, compared with CT
(35), while avoiding exposure to ionizing
radiation.
In our study, the best estimates of
PV were obtained with MFA and planimetry as performed by reader 4 when
imaging-based determinations were compared with volumes derived from pathologic specimens. Furthermore, the MFAand planimetry-determined volumes
showed substantial overlap in slopes
and CIs, suggesting MFA as an appropriate surrogate for planimetry. In addition, MFA eliminates the time required
to draw ROIs on the numerous sections
(approximately 15–20 sections) for each
case. The active shape model takes approximately 180 seconds to segment
each three-dimensional study, with a
maximum allowed time of 200 seconds.
Thus, the MFA algorithm may be a preferable clinical tool on the basis of both
accuracy and efficiency.
Reader 1 was a group of seven radiologists who prospectively and independently interpreted prostate MR imaging during their clinical interpretation
sessions. This category included measurements generated by residents in
training with limited experience in reading prostate MR images, especially with
respect to delineating the apex and base
Radiology: Volume 262: Number 1—January 2012
n
of the prostate for measuring ellipsoid
diameters. Reader 2 was an experienced
reader who retrospectively reviewed all
91 cases for gland dimensions. Though
readers 1 and 2 showed strong agreement
with one another (slope, 0.897; 95%
CI: 0.810, 0.983), the differing slopes
suggest that ellipsoid formula–derived
volume measurements have a user dependence. It is also possible that there is
dependency on the interpretive circumstance, since reader 1 PVs were generated during clinical interpretations,
whereas reader 2 PVs were generated
in a dedicated research mode, isolated
from the distractions encountered in
clinical practice. It is interesting that
the ellipsoid formula PVs showed a
stronger relationship to planimetry PVs
than to pathologically determined PVs,
whereas MFA PVs showed a stronger
relationship to pathologically determined PVs. When the ellipsoid formula–
derived PVs were compared with the
in vivo ground truth of planimetry PVs,
the regression line slopes were closer to
one, indicating a stronger relationship
with in vivo images than with pathology
specimens.
It is encouraging that all calculated
volume strategies performed reasonably
well. There seem to be reasonable choices
available to the radiologist when considering volume determinations. Since
the reference standard for comparison
radiology.rsna.org
Bulman et al
is uncertain (32), it remains a choice
whether to prefer a method that aligns
better with pathologic or in vivo planimetry findings.
The correlation between results speaks
to the overall precision of volume estimation methods. However, the lower
bound slope of the best-performing volume estimation method (MFA) versus
the pathologic reference standard can
be as low as 0.800, indicating up to a
20% underestimation of PV. All methods
showed a similar and consistent underestimation of PV when compared with
the pathologic reference standard. The
clinical implications of this underestimation and inaccuracy should be further investigated. On the other hand,
it may be useful to use the equations
determined in our study on a prospective series to see whether the attendant
adjustments yield improved accuracy.
We performed our imaging comparisons as part of our routine approach to
prostate MR imaging with an endorectal
coil. We recognize that there may be
volume changes that occur in vivo in MR
imaging with an endorectal coil; however, there are also potential volume
changes in the ex vivo specimens that
occur prior to weight measurement. Both
of these may be implicated in the increased strength of association between
MR imaging ellipsoid formula–generated
PVs and planimetry-generated PVs.
A surrogate in vivo ground truth,
planimetry, was used in our study to
approximate changes that may occur ex
vivo and for comparison with the volume estimates derived from MR images
with an endorectal coil. Heijmink et al
(36) showed a significant difference in
PV (mean 18% decrease) when MR
images were acquired with an endorectal coil compared with those acquired
with a body-array coil. In that study,
the anteroposterior dimension was, on
average, reduced by 15.7% with an endorectal coil. Such shape and volume
changes may be influenced by water loss
or physical vasoconstriction from the
pressure of the inflated endorectal balloon (36). Our results show an average
underestimation across the five volume
measurements of 15.8% when compared with prostatectomy specimens,
149
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
which is consistent with the findings of
Heijmink et al. As a point in favor of
imaging with an endorectal coil, the
signal-to-noise benefit enables higher
spatial resolution for a given imaging
time and, therefore, may yield a more
accurate delineation of gland borders.
Furthermore, most studies indicate that
the use of an endorectal coil has diagnostic advantages (27,28).
The small differences in MR imaging planimetry results between readers
3 and 4 may be related to experience in
determining the prostate gland borders.
The results of reader 3 likely benefited
from the fact that reader 4 designated
the inferior-most and superior-most
sections for reader 3. In general, inaccuracies in planimetry may be related
to volume averaging and difficulty in delineating prostate borders, especially at
the apex.
The segmentation algorithm used in
our study is fully automated and efficient with use of MFAs. To optimize
precision in volume determinations, the
MFA technology must be trained. An
important aspect of our study was that
training was a one-time process, as opposed to other studies in which training
is an iterative process done multiple
times. In our data, we trained the MFA
on a series of T2-weighted MR prostate
images, but it should be noted that this
method is generalizable to other organs
or modalities, provided that a sufficiently
representative training population is available for generating accurate statistical
shape and appearance models for the
MFA. A clear application would be in
the kidney, where distinct boundaries
and image contrast exist between the
retroperitoneal fat and the soft tissue
of the renal parenchyma.
In our study, the algorithm used to
generate test set volumes was trained
on 31 images that were not included in
the test set. With smaller training set
sizes (like our training set), it is critically important to include a sample that
employs a wide range of prostate shapes
and sizes (mean PV, 52.1 mL; range,
25–112 mL). While improved performance of the algorithm is expected to
result from employing a larger training
set, it would be helpful to investigate
150
Bulman et al
the effects of training set content and
size on algorithm accuracy.
We subtracted 3.8 g from each
specimen weight to adjust for seminal
vesicle contributions to PV. This produced a shift in data and did not affect
the correlation. Despite the strong correlation between specimen weight and
true volume, seminal vesicle size can
vary substantially between patients, resulting in a small source of error in our
study. We recognize concerns about the
use of prostatectomy specimens as the
reference standard considering the potential ex vivo blood loss and the inclusion of periprostatic tissue (37). Future
prospective studies with more precise
pathologic analysis (removal of seminal
vesicles and periprostatic tissue, immediate weight measurement) may yield
additional information on the accuracy
of in vivo volume estimates with MFA
and MR imaging.
Additionally, it would be useful to
investigate a larger data set that included
a greater number of glands, both smaller
and larger, than in our series (smaller
than 30 mL and larger than 60 mL)
to assess the accuracy of MFA-derived
PVs specifically in these outliers, where
other methods have frequently fallen
short.
We note that no significant differences between methods were demonstrated; therefore, it may require larger
sample sizes to reveal such differences.
However, the trend points toward MFA
as having the slope closest to one compared with the pathologic reference standard; therefore, it seems reasonable that
MFA can perform at least as well as
other methods and should be considered on the basis of its expected practicality in clinical practice. While our
implementation of this MFA is not available commercially, our study points toward the feasibility of an automated approach for in vivo PV determinations.
In conclusion, PV estimates with MFAs
on axial T2-weighted MR images yield
strong approximations of prostatectomy
specimen–determined PVs and can serve
as a surrogate for MR imaging planimetry determined PVs, offering the prospect for accurate volume determinations
in clinical practice.
Disclosures of Potential Conflicts of Interest:
J.C.B. No potential conflicts of interest to disclose. R.T. No potential conflicts of interest to
disclose. A.D.P. No potential conflicts of interest
to disclose. B.N.B. Financial activities related to
the present article: none to disclose. Financial
activities not related to the present article: is
a consultant for Medrad and has provided expert testimony for Medrad. Other relationships:
none to disclose. C.J.M. No potential conflicts of
interest to disclose. L.N. No potential conflicts
of interest to disclose. A.M. Financial activities
related to the present article: none to disclose.
Financial activities not related to the present article: holds various patents; is the president and
cofounder of Ibris; is the majority stock holder
in Ibris and Vascuvis . Other relationships: none
to disclose. N.M.R. No potential conflicts of interest to disclose.
References
1. 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(6):
1565–1570.
2. 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(5):1536–1540.
3. Raventós CX, Orsola A, de Torres I, et al.
Preoperative prediction of pathologically insignificant prostate cancer in radical prostatectomy specimens: the role of prostate
volume and the number of positive cores.
Urol Int 2010;84(2):153–158.
4. Matthews GJ, Motta J, Fracehia JA. The
accuracy of transrectal ultrasound prostate
volume estimation: clinical correlations.
J Clin Ultrasound 1996;24(9):501–505.
5. Al-Qaisieh B, Ash D, Bottomley DM, Carey
BM. Impact of prostate volume evaluation
by different observers on CT-based postimplant dosimetry. Radiother Oncol 2002;
62(3):267–273.
6. Aizer AA, Anderson NS, Oh SC, et al. The
impact of pretreatment prostate volume on
severe acute genitourinary toxicity in prostate cancer patients treated with intensitymodulated radiation therapy. Int J Radiat
Oncol Biol Phys 2011;79(2):379–384.
7. Caso JR, Tsivian M, Mouraviev V, Polascik TJ.
Predicting biopsy-proven prostate cancer
recurrence following cryosurgery. Urol Oncol doi:10.1016/j.urolonc.2010.04.001. Published online September 6, 2010. Accessed
December 28, 2010.
radiology.rsna.org
n
Radiology: Volume 262: Number 1—January 2012
GENITOURINARY IMAGING: Computer-derived Prostate Volumes from MR Imaging Data
8. 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(1):18–29.
9. Rodríguez Alonso A, González Blanco A,
Barbagelata López A, et al. Importance of
PSA velocity and PSA density in the prediction of prostate cancer in TURP or open
prostatectomy specimen of patients with previous negative prostate biopsy [in Spanish].
Actas Urol Esp 2008;32(8):779–786.
10. Choi YJ, Kim JK, Kim HJ, Cho KS. Interobserver variability of transrectal ultrasound
for prostate volume measurement according
to volume and observer experience. AJR
Am J Roentgenol 2009;192(2):444–449.
11. Bazinet M, Karakiewicz PI, Aprikian AG,
et al. Reassessment of nonplanimetric transrectal ultrasound prostate volume estimates.
Urology 1996;47(6):857–862.
12. 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(4):323–327.
13. 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(2):179–185.
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(8):
846–853.
15. Eri LM, Thomassen H, Brennhovd B, Håheim
LL. Accuracy and repeatability of prostate
volume measurements by transrectal ultrasound. Prostate Cancer Prostatic Dis 2002;
5(4):273–278.
16. Kimura A, Kurooka Y, Kitamura T, Kawabe K.
Biplane planimetry as a new method for prostatic volume calculation in transrectal ultrasonography. Int J Urol 1997;4(2):152–156.
17. MacMahon PJ, Kennedy AM, Murphy DT,
Maher M, McNicholas MM. Modified prostate
volume algorithm improves transrectal US
volume estimation in men presenting for
Radiology: Volume 262: Number 1—January 2012
n
prostate brachytherapy. Radiology 2009;
250(1):273–280.
18. Giubilei G, Ponchietti R, Biscioni S, et al.
Accuracy of prostate volume measurements using transrectal multiplanar threedimensional sonography. Int J Urol 2005;
12(10):936–938.
19. Nunez-Nateras R, Andrews JR, Martin GL,
et al. Accuracy of ultrasound in estimation
of prostate weight: comparison of urologists
and radiologists. Can J Urol 2010;17(1):
4985–4988.
20. Toth R, Tiwari P, Rosen M, et al. A magnetic resonance spectroscopy driven initialization scheme for active shape model based
prostate segmentation. Med Image Anal
2011;15(2):214–225.
21. Tiwari P, Rosen M, Madabhushi A. A hierarchical spectral clustering and nonlinear
dimensionality reduction scheme for detection of prostate cancer from magnetic resonance spectroscopy (MRS). Med Phys 2009;
36(9):3927–3939.
22. Monaco JP, Tomaszewski JE, Feldman MD,
et al. High-throughput detection of prostate
cancer in histological sections using probabilistic pairwise Markov models. Med Image
Anal 2010;14(4):617–629.
23. Madabhushi A, Feldman MD, Metaxas DN,
Tomaszeweski J, Chute D. Automated detection of prostatic adenocarcinoma from
high-resolution ex vivo MRI. IEEE Trans
Med Imaging 2005;24(12):1611–1625.
24. 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(2):501–503.
25. Varma M, Morgan JM. The weight of the prostate gland is an excellent surrogate for gland
volume. Histopathology 2010;57(1):55–58.
Bulman et al
28. Wang L, Mullerad M, Chen HN, et al.
Prostate cancer: incremental value of endorectal MR imaging findings for prediction of extracapsular extension. Radiology
2004;232(1):133–139.
29. Rosen Y, Bloch BN, Lenkinski RE, Greenman RL, Marquis RP, Rofsky NM. 3T MR
of the prostate: reducing susceptibility gradients by inflating the endorectal coil with a
barium sulfate suspension. Magn Reson Med
2007;57(5):898–904.
30. Cootes TF, Taylor CJ, Cooper DH, Graham J.
Active shape models: their training and application. Comput Vis Image Underst 1995;
61(1):38–59.
31. Toth R, Doyle S, Pungavkar S, Kalyanpur A,
Madabhushi A. A boosted ensemble scheme
for accurate landmark detection for active
shape models. Proceedings of the IEEE International Symposium on Biomedical Imaging: From Nano to Macro. Rotterdam, the
Netherlands: IEEE, 2010; 1081–1084.
32. Toth R, Bloch BN, Genega EM, et al. Accurate prostate volume estimation using multifeature active shape models on T2-weighted
MRI. Acad Radiol 2011;18(6):745–754.
33. Laird NM, Ware JH. Random-effects
models for longitudinal data. Biometrics
1982;38(4):963–974.
34. Milosevic M, Voruganti S, Blend R, et al.
Magnetic resonance imaging (MRI) for localization of the prostatic apex: comparison to
computed tomography (CT) and urethrography. Radiother Oncol 1998;47(3):277–284.
35. Dubois DF, Prestidge BR, Hotchkiss LA,
Prete JJ, Bice WS Jr. Intraobserver and interobserver variability of MR imaging- and CTderived prostate volumes after transperineal
interstitial permanent prostate brachytherapy.
Radiology 1998;207(3):785–789.
26. Ohlsén H, Ekman P, Ringertz H. Assessment
of prostatic size with computed tomography:
methodologic aspects. Acta Radiol Diagn
(Stockh) 1982;23(3A):219–223.
36. 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(5):1446–1453.
27. Masterson TA, Touijer K. The role of endorectal coil MRI in preoperative staging
and decision-making for the treatment of
clinically localized prostate cancer. MAGMA
2008;21(6):371–377.
37. 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(6):935–940.
radiology.rsna.org
151