Segmentation Propagation for the Automated ...

Segmentation Propagation for the Automated Quantification of Ventricle Volume from Serial MRI Marius George Linguraru and John A. Butman Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, Bethesda MD, USA

ABSTRACT Accurate ventricle volume estimates could potentially improve the understanding and diagnosis of communicating hydrocephalus. Postoperative communicating hydrocephalus has been recognized in patients with brain tumors where the changes in ventricle volume can be difficult to identify, particularly over short time intervals. Because of the complex alterations of brain morphology in these patients, the segmentation of brain ventricles is challenging. Our method evaluates ventricle size from serial brain MRI examinations; we (i) combined serial images to increase SNR, (ii) automatically segmented this image to generate a ventricle template using fast marching methods and geodesic active contours, and (iii) propagated the segmentation using deformable registration of the original MRI datasets. By applying this deformation to the ventricle template, serial volume estimates were obtained in a robust manner from routine clinical images (0.93 overlap) and their variation analyzed. Keywords: Brain imaging, serial MRI, brain tumor, hydrocephalus, segmentation, registration.

1.

INTRODUCTION

Serial brain MRI examinations are performed to monitor patients with brain lesions. The routine use of high resolution 3D imaging and coregistration facilitates such evaluation. In addition, the routine clinical observation of such coregistered studies suggests that that ventricle size progressively increases in many of these patients. This is a poorly documented phenomenon, likely representing a communicating hydrocephalus [4]. The etiology is unclear, and is possibly related to the introduction of blood and proteins into the cerebral-spinal fluid (CSF) during surgery. The change in ventricular size may be quite subtle even by comparison of coregistered images. Thus, it is desirable to identify the incidence of the phenomenon, and to correlate the presence or absence of such a phenomenon with clinical symptoms. Therefore, we sought an objective method to systematically characterize the clinical observation of progressive ventriculomegaly in this group. Because of the complex alterations of brain morphology in our patients, the segmentation of brain ventricles is challenging. T1 images obtained post contrast add the complication of enhancing tissue (the choroid plexus) in the ventricles, raising the number of tissue classes. Edema and gliosis also result in signal intensities overlapping with CSF. Surgical intervention results in communication between the ventricular system and the cisterns. These factors complicate the segmentation process. In previous work, ventricles were segmented using region-growing combined with anatomical knowledge [9] in images of high quality. In [2], non-rigid registration was combined with data from an atlas to segment brain ventricles, relying on the high performance of the registration algorithm. Because of the complex alterations of brain morphology in our patients, we do not expect atlas based segmentation to work robustly. The registration with an atlas is also addressed in [6] using an expectation-maximization joint model. Most recently, Hu and Collins [3] proposed a model-based segmentation on multi-modality MR images.

Medical Imaging 2009: Computer-Aided Diagnosis, edited by Nico Karssemeijer, Maryellen L. Giger Proc. of SPIE Vol. 7260, 72601Z · © 2009 SPIE · CCC code: 1605-7422/09/$18 · doi: 10.1117/12.810940

Proc. of SPIE Vol. 7260 72601Z-1

Many prior studies use high resolution data from uniform data sets. Because our data comes from patients with brain tumors on a number of different imaging platforms, image contrast can be quite variable and signal-to noise ratio (SNR) low. Hence, our method was designed to exploit patient-specific data for better-adapted intra-patient non-rigid registration, less sensitive to image quality and the anatomical variability of brain tumor patients. Mean T1 images with increased SNR provide the input for the segmentation based on geodesic active contours. Hence, no correction for inhomogeneous signal intensity is required. The seeding for the segmentation is found automatically through the analysis of the CSF. Finally, the mean ventricular shape is adaptively propagated through the temporal data to quantify size changes

2. METHOD Patients with brain tumors evaluated at the Clinical Center of the NIH were scanned on 1.5 T GE (Milwaukee, WI) or 3 T Philips (Best, Netherlands) MR systems at approximately 1 to 3 month intervals. As part of the routine clinical imaging, 3D T1 weighted sequences were obtained following injection of an intravenous contrast agent. On the 1.5 T scanner a 3D SPGR technique was used with TR 12 ms; TE 5 ms; flip angle 20 degrees; 240 cm FOV; 256 matrix sagittal acquisition, 152 slices 2 mm thick with 1 mm overlap. On the 3.0 T scanner the parameters were TR 5 ms; TE 2 ms; flip angle 15 degrees, 240 cm FOV, 256 matrix; 1 mm slice thickness. 2.1 Pre- processing For each patient, the 3D datasets were coregistered to the first dataset in the series using FLIRT (FMRIB, Oxford, UK) under MEDx (Medical Numerics, Sterling, VA). Coregistration was done by a 6 parameter rigid body transformation (translation + rotation) using a least squares cost function. To normalize the dataset, a large ~100 cc3 volume of interest (VOI) was automatically placed in the brain volume in relation to its center of mass. The histogram of this VOI was used to identify the modal signal intensity, presumably representing white matter. Each dataset was normalized by dividing by this value, and a high signal-to-noise ration (SNR) “mean” image was generated for each patient by averaging the time point datasets using AFNI (http://afni.nimh.nih.gov/). Finally, smoothing was performed by anisotropic diffusion [5]. 2.2 Segmentation The level sets are initialized from an automatically set seed on the septum pellucidum (between the bodies of the lateral ventricles). A subvolume of the scan, centered on the x and y axes and large enough to include the superior lateral ventricles, is thresholded to keep the pixels in the lowest bin of the histogram after intensity normalization. The resulting binary CSF mask is morphologically closed to connect the lateral ventricles and the 3D object is decomposed in 2D axial slices. The connected components of each slice are analyzed and we select the slice with the largest CSF object, presumably at the location of the septum pellucidum. The centroid of the largest CSF object is set to be the segmentation seed. The second stage of the method is the segmentation of lateral ventricles and our approach uses a combination of fast marching and geodesic active contour level sets [1,8]. The fast marching method assumes that the surface can only expand staring from the seed point. . The speed of expansion is constant and along the surface normal n. The MRI scan I provides the feature image, while the sigmoid of the gradient of I supplies the speed function Ie. The first segmentation given by the fast marching level set is If.

dI f dt

+ nI e ∇I f = 0

A better-adapted level set based on geodesic active contours in used to refine the fast marching segmentation [1]. To initialize the model, we use the fast marching segmentation as input level image (zero-level) into the geodesic active

Proc. of SPIE Vol. 7260 72601Z-2

contour IL. The weights w1, w2 and w3 control respectively the speed c, curvature k and attraction to edges (Caselles et al. 1997).

dI L = I e (w1c + w2 k ) ∇I L + w2 ∇I e ∇I L dt The level sets are initilized from an automatically set seed on the septum pellucidum (between the bodies of the lateral ventricles). A subvolume of the scan, centered on the x and y axes and large enough to include the superior lateral ventricles, is thersholded to maintain only the pixels in the lowest bin of the histogram. This gray level corresponds to the CSF after the intensity normalization of MR scnas. The resulting binary CSF mask is morphologically closed to connect the lateral ventricles and the 3D object is decomposed in 2D axial slices. The connected components of each slice are analyzed and we select the slice with the largest CSF object, presumably at the location of the septum pellucium. The centroid of the largest CSF object is set to be the segmentation seed, as shown in Figure 1.

Figure 1. The position of the user specified seed to initiate the ventricle segmentation on the mean image. Note the intermediate intensity of the tissue adjacent to the posterior portion of t he body of the right lateral ventricle. This is due to the averaging of datasets at varying stages of ventricle dilation in this location.

2.3 Propagation A refinement of the intra-patient registration is required to compensate for the residual deformation not covered by the rigid registration used in preprocessing. Differences in the brain anatomy are mainly due to effects of therapeutic interventions, such as surgery and chemotherapy as well as disease progression (e.g. tumor growth or hydrocephalus). We propose employing the non-rigid registration algorithm based on B-splines [7]. The deformation of objects is governed by an underlying mesh of control points in coarse to fine multiresolution approach. B-splines allow to locally control the deformation T and by varying the spacing between control points, the number of degrees of freedom is adapted to account for more global (affine) or local (non-rigid) transforms. Finally, a compromise between the similarity provided by mutual information M [10] and smoothing S is searched, where p(I,J) is the joint probability distribution of images I and J, and p(I) and p(J) their marginal distributions. For more detail on the B-spline definition of the transformation T, please refer to [7]. arg min [M (I | T ( I ) ) − S (T ) ] p (I ) + p ( J ) M (I | J ) = p (I , J ) S (T ) =

∫ (∂ T ) 2

x, y ,z

dxdydz

x, y ,z

Proc. of SPIE Vol. 7260 72601Z-3

The resulting deformation fields between the mean image and the temporary acquisitions of each patient are applied to the ventricle mask segmented from the mean image using a nearest-neighbor interpolation. The deformations are intrapatient and well defined, and provide a robust propagation of the segmentation. These processes were implemented under Visual C++ 8.0 (Microsoft), OpenGL (SGI) and the Insight Segmentation and Registration Toolkit (ITK) 3.4.

3

RESULTS

Ventricle size was analyzed from sixteen patients who were imaged with MRI between 8 and 25 times at 1 to 14 month intervals over a minimum of 17 and a maximum of 53 months. The total number of time-point scans was 201. The results of the semi-automated assessment method of ventricles from the T1 brain MRI are shown in Figure 2. The MRI represents the mean image resulting from rigid transformation, intensity normalization and averaging. This mean image has increased the signal-to-noise ratio (SNR) as compared to a single scans (Figure 3). Subsequently, the lateral ventricles were segmented from the mean data, using fast marching and geodesic active contour level sets, as seen in Figure 2.

Figure 2. The 3D segmentation of brain ventricles from the mean image. Note the improved signal-to-noise-ratio (SNR) of the mean image, compared to SNR of data in Figures 3 and 4.

The automatic initialization of the seed failed in 3 of the 16 cases, where it segmented only one of the lateral ventricles. Two of the cases had very narrow ventricles, while the third case had high vascularizition caused by the vicinity of a tumor. In these cases, the segmentation was initialized manually by placing the seed. In these difficult cases, the segmentation of the posterior lateral ventricles is not accurate; nevertheless, the trend of ventricular volume change is estimated correctly. To obtain the segmentation of ventricles on each scan of the time series, the mean image was registered to each individual scan of the same patient by non-rigid registration and the deformation fields were saved. This deformation field was then applied to the ventricle mask, thereby propagating the segmentation through each individual time point of the series. The average image may not conform to any individual dataset, but it does to a greater degree than an atlas from another subject. The deformation results in a good match to each individual dataset. An example of the robustness of the 3D segmentation of ventricles across the entire series of a patient’s temporal acquisition is shown in Figure 3. Results in the axial plane of temporal scans of a different patient are then presented in Figure 3.

Proc. of SPIE Vol. 7260 72601Z-4

Figure 3. The 3D segmentation of brain ventricles from a single time point of a patient.

Figure 4. 21 rigidly coregistered MRI sets obtained at 1-3 month intervals demonstrating the change in ventricle size over time, and the segmentation obtained. The initial MRI is top left, and the final MRI is Bottom right. The segmentation results correlate well with the steady increase in ventricular volume and change of ventricular shape. Note that there is considerable variation in image contrast due to changes in tumor and differences in scanners used.

For the quantification of segmentation/propagation results, we compared the data processed by our algorithm with manually-segmented data. A total of 15 individual MRI scans from 3 patients were analyzed (5 acquisitions/patient) and the manually measured volumes of the lateral ventricles were recorded. Figure 5 presents the temporal evolution of the automatically evaluated ventricular changes near the manually measured volumes (in mm3). Although the change in size

Proc. of SPIE Vol. 7260 72601Z-5

between subsequent 3D ventricular bodies is not obvious for the eye, the volume of lateral ventricles changes substantially, as measure by our method and seen in Figures 5 and 6. The mean error in volume estimation is of 3.58+/3.66%. The DICE overlap between manual and automated segmentations was 0.93+/-0.01. Evaluation of ventricular volume over time in this non random sample showed ventricle volume to be stable over time in 3 cases, slowly increasing in 11 cases, rapidly increasing in one case (due to obstructive hydrocephalus), and decreasing in one case (due to mass effect of an enlarging tumor). This concurred with the visual interpretation of the coregistered datasets by an experienced radiologist.

90000 800Uu

70000 60000 -

-

E

--'

50000 4000U

> 30000 A

"itt' 20000 -

Automated

A ManuI

10000 U

0

5

10

15

2fl

25

30

35

40

Months

Figure 5. Manual versus automated ventricular size estimation. This example shows the automated quantification of ventricular growth in a brain tumor patient (same as in Figure 2), along the manual measurements of ventricular volume.

Figure 6. The 3D automated ventricular size estimation at the points of comparison with manual segmentation (blue point in Figure 3). Although the change in size between subsequent 3D ventricular bodies is not always obvious for the eye, the volume of lateral ventricles changes substantially, as measured by our method and seen in Figure 3.

Proc. of SPIE Vol. 7260 72601Z-6

4

DISCUSSION

The segmentation results are robust throughout the database and the segmentation propagation based on the deformation fields accurate, as confirmed by experienced radiologists. In this small sample, ventricle volume was stable or increases in size relatively slowly and uniformly in many cases. Departures from this time course could be related to changes in the clinical picture (e.g. development of obstruction). These initial results show great promise toward an automated reliable tool for ventricular size change assessment and confirmation of radiological observations related to the possible relation between brain tumor resection and communicating hydrocephalus. Segmenting the ventricles is relatively easy in healthy volunteers and patients with uniform disease processes. Also, the rate of growth of a uniform expansion of the ventricular system could be modeled with a coarse or even a partial segmentation. However, our method addresses the uniquely difficult data of patients with brain tumors and surgical interventions, which result in local deformations of the ventricles. T1 images obtained post contrast add the complication of enhancing tissue in the ventricles, raising the number of tissue classes. Other factors that hamper the segmentation process are the presence of edema and gliosis with signal intensities overlapping with CSF, and surgical intervention resulting in communication between the ventricular system and the cisterns. The segmentation of an individual dataset is also unreliable in this noisy data; hence we were unable to compare propagating the deformable model with serial segmentation done by the same algorithm, but the overlap with manually segmented data was of 93%. Moreover, clinical exigencies preclude scanning the same patient on the same scanner with the same sequence over time. This may be possible in very restrictive research situations, but is not generally possible, making serial segmentation more likely to inform us about differences in scan techniques across time. There are several advantages to this technique. Because the segmentation is initially performed on a high SNR dataset is used to define the ventricles in all the component data sets, it is not necessary for the segmentation to be successful on each individual dataset. Therefore an individual dataset which is noisy or artifact ridden (e.g. from patient motion) can be evaluated by segmentation propagation. Furthermore, disruption of the ventricle border due to surgical intervention or tumor growth may cause region growing or level set segmentation to “leak out” of the ventricular system. In this method, this needs only be dealt with on the mean data set and not on each individual data set. The drawbacks of the technique come from the time-consuming non-linear registration and the errors in segmentation around the contrast-enhanced blood vessels in the ventricles, which can block the expansion of the levet sets.The optimization of the program along with the use of more powerful processing units are currently under investigation. A larger database and an extended evaluation of the method will follow. The algorithm will allow the distinction between cases that show an increase in ventricular size and those without hydrocephalus. Further development will lead to a fully-automated assessment method ready to be used in daily clinical practice and support the documentation of the tumor related brain atrophy. Because our data comes from patients with brain tumors on a number of different imaging platforms, our method was designed to exploit patient-specific data for better-adapted intra-patient non-rigid registration, less sensitive to image quality and the anatomical variability of brain tumor patients. Mean T1 images with increased SNR provide the input for the segmentation based on geodesic active contours. Hence, no correction for inhomogeneous signal intensity is required. The seeding for the segmentation is found automatically through the analysis of the CSF. Finally, the mean ventricular shape is adaptively propagated through the temporal data to quantify size changes. Also, the method allows for the first time to monitor patients with brain lesions toward documenting the possible communicating hydrocephalus phenomenon in this category of patients.

Proc. of SPIE Vol. 7260 72601Z-7

5

CONCLUSION

Results show great promise toward an automated reliable tool for ventricular size change assessment; in this particular case the confirmation of radiological observations related to the possible relation between brain tumor resection and communicating hydrocephalus. The algorithm will allow the distinction between cases that show an increase in ventricular size and those without hydrocephalus for improved clinical management of brain diseases. Acknowledgment This work was supported by the Intramural Research Program of the National Institutes of Health, Clinical Center.

REFERENCES [1] Caselles, V. Kimmel, R. Sapiro, G., “Geodesic active contours,” International Journal on Computer Vision 22(1), 61–97 (1997). [2] Holden, M. Schnabel, J.A. Hill, D.L. “Quantification of small cerebral ventricular volume changes in treated growth hormone patients using nonrigid registration,” IEEE Transactions on Medical Imaging 21(10), 1292-1301 (2002). [3] Hu, S. Collins, L., “Joint level-set shape modeling and appearance modeling for brain structure segmentation,” Neuroimage 36(3), 672-683 (2007). [4] Inamasu, J. Nakamura, R. Saito, Y. Kuroshima, Y. Mayanagi, K. Orii, M. Ichikizaki K., “Postoperative communicating hydrocephalus in patients with supratentorial malignant glioma,” Clinical Neurology and Neurosurgery 106(1), 9-15 (2003). [5] Perona, P. Malik, J., “Scale-space and edge detection using anisotropic diffusion,” IEEE Transactions on Pattern Analysis and Machine Intelligence 12, 629–639 (1990). [6] Pohl, K.M. Fisher, J. Grimson, W.E. Kikinis, R. Wells, W.M., “A Bayesian model for joint segmentation and registration,” Neuroimage 31(1), 228-239 (2006). [7] Rueckert, D. et al., “Non-rigid registration using free-form deformations: Application to breast MR images,” IEEE Transactions on Medical Imaging 18(8), 712-21 (1999) [8] Sethian, J.A., [Level Set Methods and Fast Marching Methods], Cambridge University Press, (1996). [9] Schnack, H.G. Hulshoff, H.E. Baaré, W.F. Viergever M.A. Kahn, R.S., “Automatic segmentation of the ventricular system from MR images of the human brain,” Neuroimage 14(1), 95-104 (2001). [10] Studholme, C. Hill, D.L.G. Hawkes, D.J., “An overlap invariant entropy measure of 3D medical image alignment,” Pattern Recognition 32(1), 71-86 (1999).

Proc. of SPIE Vol. 7260 72601Z-8