## Accurate prostate volume estimation using multifeature active shape models on T2-weighted MRI

Venue: | Acad. Radiol |

Citations: | 1 - 0 self |

### BibTeX

@ARTICLE{Toth_accurateprostate,

author = {MS Robert Toth and MD B Nicolas Bloch and MD Elizabeth M Genega and MD Neil M Rofsky and PhD Robert E Lenkinski and MD Mark A Rosen and PhD Arjun Kalyanpur and MD Sona Pungavkar and MD Anant Madabhushi},

title = {Accurate prostate volume estimation using multifeature active shape models on T2-weighted MRI},

journal = {Acad. Radiol},

year = {},

pages = {745--754}

}

### OpenURL

### Abstract

Rationale and Objectives: Accurate prostate volume estimation is useful for calculating prostate-specific antigen density and in evaluating posttreatment response. In the clinic, prostate volume estimation involves modeling the prostate as an ellipsoid or a spheroid from transrectal ultrasound, or T2-weighted magnetic resonance imaging (MRI). However, this requires some degree of manual intervention, and may not always yield accurate estimates. In this article, we present a multifeature active shape model (MFA) based segmentation scheme for estimating prostate volume from in vivo T2-weighted MRI. Materials and Methods: We aim to automatically determine the location of the prostate boundary on in vivo T2-weighted MRI, and subsequently determine the area of the prostate on each slice. The resulting planimetric areas are aggregated to yield the volume of the prostate for a given patient. Using a set of training images, the MFA learns the most discriminating statistical texture descriptors of the prostate boundary via a forward feature selection algorithm. After identification of the optimal image features, the MFA is deformed to accurately fit the prostate border. An expert radiologist segmented the prostate boundary on each slice and the planimetric aggregation of the enclosed areas yielded the ground truth prostate volume estimate. The volume estimation obtained via the MFA was then compared against volume estimations obtained via the ellipsoidal, Myschetzky, and prolated spheroids models. Results: We evaluated our MFA volume estimation method on a total 45 T2-weighted in vivo MRI studies, corresponding to both 1.5 Tesla and 3.0 Tesla field strengths. The results revealed that the ellipsoidal, Myschetzky, and prolate spheroid models overestimated prostate volumes, with volume fractions of 1.14, 1.53, and 1.96, respectively. By comparison, the MFA yielded a mean volume fraction of 1.05, evaluated using a fivefold cross-validation scheme. A correlation with the ground truth volume estimations showed that the MFA had an r 2 value of 0.82, whereas the clinical volume estimation schemes had a maximum value of 0.70. Conclusions: Our MFA scheme involves minimal user intervention, is computationally efficient and results in volume estimations more accurate than state of the art clinical models. Key Words: Prostate volume; active shape models; prostate cancer; MRI; texture; image processing. ªAUR, 2011 P rostate volume has been shown to be a strong predictor of treatment outcome for patients with prostate cancer (1,2), especially when combined with a baseline prostate-specific antigen (PSA) level (3). Prostate volume has also been shown to be useful in determining PSA density (4). The most common method for estimating the prostate volume involves modeling the prostate as a simple geometric shape based on manually estimated measurements of the anteroposterior, transverse, and craniocaudal lengths of the prostate. The most common models for approximating the prostate shape are the ellipsoid model 745 Although most prostate volume estimations are done using TRUS imagery, a strong correlation (r 2 = 0.925) has been shown between the volume estimations obtained using TRUS and from magnetic resonance imaging (MRI) (5). In addition, the ellipsoidal model was found to yield accurate volume estimations for T2-weighted MRI of the prostate, even when an endorectal coil was used (12). In previous work (13), it was found that the ellipsoidal volume estimations were more accurate than a planimetry-based approach (aggregating a series of measurements from each slice) when using a surface coil; in contrast to other work (12) in which planimetry estimates were found to yield more accurate volume estimations compared to the ellipsoidal model estimates when using an endorectal coil. In previous work (5), a planimetry based volume estimation was performed by measuring the areas from manual two-dimensional (2D) segmentations of the prostate on each slice. Our prostate volume estimation method is related to the technique used by Hoffelt et al Although active shape models (ASMs) are a popular segmentation technique, they sometimes fail to converge to the desired object boundary in the case of weak image gradients (19). ASMs essentially model the shape of an object a statistical variations in a set of anatomical landmarks the appearance of an object as a Gaussian distribution of intensities near each anatomical landmark. The appearance model typically uses the intensities of the image to learn a statistical appearance model. However, relying solely on the intensity information may not be sufficient for accurately detecting the correct boundary, especially if different regions of the image, or different regions within the desired object, have similar intensity values. This is particularly true of MRI in which strong bias field inhomogeneity artifacts can significantly obfuscate object boundaries In this work we present a new ASM that we call the MFA. We calculate the gray level statistics of each image by convolving a set of kernels with the intensity image. These include the Kirsch (21) and Sobel (22) kernels to better quantify the edges of the prostate border. Although traditional ASMs use neighboring intensity information, they are dependent on the normal to the shape at any given landmark point. By contrast, the Gaussian and mean kernels take neighboring information into account and yet do not depend on the normal of the shape. Additionally, the Cartesian x and y coordinates of each landmark point are included as additional ''features.'' Further, because texture features of the prostate boundary are not always optimally modeled as a Gaussian, we describe the distributions as sums of multiple Gaussians (GMM) (23), allowing us to better characterize the feature distributions at each landmark on the prostate boundary. A forward feature selection scheme is employed to determine the best textural features in terms of discriminability between the prostate border and background. Only these features are then employed in conjunction with the MFA. The MFA is employed to estimate the gland area on each slice, which is multiplied by the slice interval (distance between center of adjacent slices) to yield an estimation of the prostate volume. This estimation is compared to the ellipsoid (4), Myschetzky (15), and prolate spheroid (4) volume estimation techniques. All four methods were evaluated in terms of accuracy with respect to a ground truth estimate of the prostate volume obtained via expert radiologist derived segmentations of the prostate on individual 2D slices. MATERIALS AND METHODS Data Description and Notation The datasets considered in this study comprised 19 1.5 Tesla (T) MRI studies obtained from the American College Of Radiology Imaging Network trial (24) and 26 3T T2-weighted MRI studies from the Beth Israel Medical Center in Boston, henceforth denoted as D 1 and D 2 respectively. A complete description of the 45 MRI datasets considered in this study is provided in Ground Truth Estimations of Prostate Volume The ground truth volume (V Ex ) for the prostate in each of the 45 studies was determined as follows. For each study C, an expert radiologist provided a manual segmentation of the prostate for all slices in which the prostate was visible. The set comprising the area estimates of the prostate from all M slices within a single three-dimensional (3D) study C, is denoted as S Ex = {A m , j m˛{1, . . ., M}} where A m denotes the segmented area of 2D slice m. The estimated prostate areas (region contained within the manual delineations of the capsule) on all slices are integrated and multiplied by the slice interval T. This is similar to the approach presented elsewhere (5), in which planimetry area estimates were aggregated to estimate the prostate volume. The ground truth prostate volume (V Ex ) in C is then calculated as (1) Clinically Employed Prostate Volume Estimation Models For the ellipsoid, Myschetzky, and prolate spheroid models, an expert manually determined the transverse (D 1 ), craniocaudal MFA-based Prostate Volume Estimations (V MFA ) The MFA is a novel extension of the traditional ASM (18), but uses multiple texture features to characterize the prostate border. The MFA contribution comprises the 5 main steps, which are summarized in Generating a statistical shape model. For each slice from each training image, 100 landmarks are manually placed along the prostate border. X is used to represent a series of 100 x and y Cartesian coordinates, so that X˛ℝ 200 . Principal component analysis is performed on all X (18), so that the shape of the prostate can be characterized by 10 parameters, in turn explaining 98% of the variation seen in the prostate shape. This 10-dimensional vector of parameters b˛ℝ 10 can now be employed to describe a specific shape X b . The details of this system are shown in the Appendix, which can be accessed online. Generate an appearance model. The traditional ASM methodology used neighboring intensities around each of the 100 landmark points to describe the appearance of the prostate border. In this extension, the distribution of texture features G (instead of intensities) at each landmark is modeled using a sum of multiple Gaussian distributions (23). Forward feature selection. A feed forward feature selection is employed for identification of only the most discriminatory textural attributes that are to be used in conjunction with the MFA. Thus while 50 texture features are initially generated, only a few (ie, 5) discriminating ones are selected for use in conjunction with the final appearance model. Segmentation using the MFA. Using the trained appearance model, the most probable locations of the prostate border on a new image are determined. The probability of voxel c belonging to landmark point n is denoted by P n (c). The locations b X that had the highest probability of corresponding to a landmark point were thus selected. Finally, b was modified to optimally fit b X as per the ASM fitting technique in (18), yielding a final set of landmarks X b . The landmarks X b are connected via linear interpolation, and the segmented area can then be determined. The area inside the segmentation for each slice m is given as A m . Equation 1 is then used to calculate V MFA from the segmentations of all slices in a given 3D volumetric acquisition. Experiments Performed For e 4 , a fivefold cross validation across patients was performed. To perform the cross-validation, 4/5 of the studies were randomly selected, and used to train an MFA. Then, Prolate spheroid Academic Radiology, Vol 18, No 6, June 2011 ACCURATE PROSTATE VOLUME ESTIMATION 747 the remaining 1/5 of the studies (which were not used to train) were segmented using the trained MFA. This was repeated until all studies had been tested. Thus the same studies were never used to train and test simultaneously. It should be noted that due to the extreme differences in image quality and structural information, separate cross validations were performed for D 1 and D 2 , respectively. RESULTS Pearson's Correlation Coefficient We first compared V MFA with the clinical models V Ell , V Mys , and V Sph for the 45 volumetric studies. This was done by calculating the Pearson correlation coefficient (25) (the r 2 value) between each of V MFA , V Ell , V Mys , and V Sph with V Ex over all 45 testing studies. The hypothesis for these experiments was that V MFA should have at least as high an r 2 value as V Ell , V Mys , and V Sph with respect to V Ex . The results of these experiments are shown in Comparison of Volume Fractions The volume fraction between V and V Ex was calculated for each of the 45 studies in which a value of 1.00 indicates that the estimated volume is exactly equal to the ground truth volume. The results from these calculations are shown in The MFA (e 4 ) had a volume fraction of 1.05 with a standard deviation of 0.21, and is shown in Statistical Significance Between Volume Fractions The MFA had a mean volume fraction V MFA /V Ex closest to 1.00, and a paired Student t-test was performed to determine if this was statistically better than each of the other volume fractions (V Ell /V Ex , V Mys /V Ex , V Sph /V Ex ). The null hypothesis, therefore, was that the mean volume fractions of the MFA (V MFA /V Ex ) and the other methodologies were equal. The results shown in DISCUSSION An automatic and reproducible method for estimating the volume of the prostate from in vivo T2-weighted MRI data using MFA. The MFA incorporated multiple statistical texture features including the Kirsch, Sobel, Gaussian, and 748 mean intensity kernels to better characterize the prostate border. In addition, GMMs were used to model the distribution of texture features instead of a traditional single Gaussian, and a forward feature selection algorithm only retained the optimal features for prostate segmentation. The MFA-based segmentation scheme had a higher correlation with the ground truth (r 2 = 0.82) compared to such traditional schemes as the ellipsoid (r 2 = 0.70), Myschetzky (r 2 = 0.70), and prolate spheroid (r 2 = 0.45) models. It was to be expected that the ellipsoidal and Myschetzky have the same r 2 value because they are simply scaled variants of each other. In addition, the prolate spheroid expectedly performed the worst of the clinical estimation techniques, as it only used two axes in its volume estimation while the ellipsoidal and Myschetzky used measurements from three axes. Qualitative results also revealed that our MFA was able to easily out-perform the traditional, intensity-driven ASM. Reasons for this include the use of textural features such as image gradients, which are not prone to intensity artifacts such as bias field. In addition, the distribution of features is not necessarily best modeled as a Gaussian, which the traditional ASM assumes. In the MFA model, we use GMMs to model our distribution of features, which can capture non-Gaussian shapes of distributions. Finally, we only retain the optimal features in our appearance model, thus automatically discounting noisy and nondiscriminatory features. That the ellipsoidal estimation performed better than the Myschetzky estimation in terms of volume fractions was not surprising because the Myschetzky correction aims to increase the ellipsoidal model's volume estimation. This would only be useful if the ellipsoidal volume estimation happened to underestimate the capsule's volume. However, unlike has been previously reported in the literature for TRUS imagery In addition, Our 2D MFA was used to generate a segmentation of the visible gland on each slice of a 3D dataset. Because the MFA models the object border using a multidimensional distribution, a large number of training images are required for accurate model generation. A 2D MFA is employed on account of the limited number of 3D studies would prevent accurate statistical models from being generated in 3D (26). However, although we only had access to a limited number 750 of 3D studies, they constituted a total of 690 2D image slices, which was more than sufficient to generate accurate statistical models in 2D (26). That the geometric based models V Ell , V Mys , V Sph performed significantly worse compared to the MFA based volume estimations V MFA was most likely because of two factors. The first potential cause is that the very tip of the base and the very tip of the apex of the prostate were used to estimate the geometric models (D 2 includes this), whereas the expert segmentations may not have necessarily included the extreme tips of the prostate. This would have yielded much higher volume estimations from the geometric models compared to the surrogate ground truth volume estimations. The second possible reason is that the geometric models are inherently convex, whereas the prostate may have distinctly concave regions. Although the ASM would be able to model these non-convex regions We also evaluated our MFA scheme in terms of segmentation accuracy by calculating the Dice similarity coefficient (28). The MFA achieved a mean Dice similarity coefficient of 0.8483 with a standard deviation of 0.0448 (standard error 0.0060) over 45 studies. This compared favorably against the prostate MRI segmentation models in Limitations of this study include the fact that expert segmentations were used as a surrogate of the ground truth volume. An alternative would have been to use the volume of the excised prostatectomy specimen as the gold standard, but these were not available for this study. A secondary, minor limitation was the limited number of 3D studies (preventing the use of a full 3D ASM). In summary, the MFA volume estimation method can save valuable time for clinicians and can yield a consistently accurate, near realtime prostate volume estimation that is extremely useful for evaluating posttreatment response. ACKNOWLEDGMENTS