Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

Automated Segmentation of MRI of Brain Tumors Michael R Kaus1,3, PhD * Simon K Warfield1, PhD * Arya Nabavi1,2, MD Peter M Black2, MD, PhD * Ferenc A Jolesz1, MD * Ron Kikinis1, MD 1

Surgical Planning Laboratory, Department of Radiology Department of Neurosurgery, Brigham & Women's Hospital, Harvard Medical School, 75 Francis Street, Boston, MA 02115 3 Lehrstuhl Technische Elektronik, University Erlangen-Nürnberg, Cauerstr. 9, 2

D - 91058 Erlangen, German Grant Support: This work was supported (in part) by a grant from the Deutscher Akademischer Austauschdienst (DAAD), (in part) by a Grant from the National Multiple Sclerosis Society (SW), (in part) by NIH grants RO1 CA 46627-08, PO1 CA67165-01A1, PO1 AG04953-14, NSF grant BES 9631710 and Darpa grant F41624-96-2-0001. AN was supported by DFG NA 359/1-1.

Abstract An automated brain tumor segmentation method was developed and validated against manual segmentation on 3D-MRI of 20 patients with meningiomas and low grade gliomas. The automated method allows the rapid identification (5-10 minutes operator time) of brain and tumor tissue with accuracy and reproducibility comparable to manual segmentation (3-5 hours operator time) making automated segmentation practical for low grade gliomas and meningiomas. Key words: Brain neoplasms, Magnetic resonance (MR), Computer assisted neurosurgery, Image segmentation

1 Introduction Computer assisted surgical planning and advanced image-guided technology have become increasingly utilized in neurosurgery [1, 2, 3, 4, 5]. The availability of accurate anatomical three dimensional (3D) models significantly improves spatial information concerning relationships of critical structures (e.g. functionally significant cortical areas, vascular structures) and pathology [6, 3, 4]. In daily clinical practice, however, commercially available intraoperative navigational systems only provide the surgeon with 2D cross-sections of the intensity value images and a 3D model of the skin. The main limiting factor in the 1 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

routine use of 3D models to identify (segment) important structures is the amount of time and effort that a trained operator has to spend on the preparation of the data [3, 6]. The development of automated segmentation methods has the potential to significantly reduce the time for this process and make such methods practical. Although 2D images accurately describe the size and location of anatomical objects, the process of generating 3D views to visualize structural information and spatial anatomic relationships is a difficult task and usually carried out in the clinician's mind. Image processing tools provide the surgeon with interactively displayed 3D visual information which is somewhat similar to the view of the surgeon during surgery, thus facilitating the comprehension of the entire anatomy. For example, the (mental) 3D visualization of structures that do not readily align with the planes of the image acquisition (e.g. the vascular tree) is extremely difficult if it is based on 2D images alone. Image based modeling requires computerized image processing methods which include segmentation, registration and display. Segmentation using statistical classification techniques [7, 8] has been successfully applied to gross tissue type identification. Because the acquisition of tissue parameters is insufficient for successful segmentation due to the lack of contrast between normal and pathologic tissue [15, 16], statistical classification may not differentiate between non-enhancing tumor and normal tissue [12, 13, 14]. Explicit anatomical information derived from a digital atlas has been used to identify normal anatomical structures [9, 10, 11]. We have developed an automated segmentation tool which can identify the skin surface, the ventricles, the brain and tumor in patients with brain neoplasms [18, 19]. The purpose of the current study was to compare the accuracy and reproducibility of this automated method with those of manual segmentation carried out by trained personnel.

2 Materials and Methods 2.1 MR Imaging Protocol Patient heads were imaged in the sagittal and axial plane with a 1.5 T MRI system (Signa, GE Medical Systems, Milwaukee, WI), with a postcontrast 3D sagittal spoiled gradient recalled (SPGR) acquisition with contiguous slices (flip angle, 45° ; repetition time (TR), 35 msec; echo time (TE), 7 msec; field of view, 240 mm; slice-thickness, 1.5 mm; 256 × 256 × 124 matrix). The acquired MR images were transferred to a UNIX network via Ethernet.

2.2 Brain Tumor Patients 20 patients were selected from a neurosurgical image database of approximately 260 brain tumor patients. Each of the 260 cases had been post-processed for image-guided neurosurgery by using a combination of semiautomated techniques and manual outlining of the skin-surface, the brain, the ventricles, the vessels and the tumor. A neurosurgeon was asked to select the 20 cases with meningiomas and low grade gliomas of different size, shape and location to provide a representative selection. These two types were selected because they are relatively homogeneous and have well defined imaging characteristics. The pathological diagnoses

2 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

included 6 meningiomas (cases No. 1-3, 11, 12, 16), and 14 low grade gliomas (cases No. 4-10, 13-15, 17-20). In this study, 6 out of 6 meningiomas enhanced well, and 14 out of 14 low grade gliomas were non-enhancing. Cases No. 1-10 formed the development database used for the design and validation of the automated segmentation method. To ensure that the method produced correct results when applied to cases other than those of the development database, in addition to the 10 development cases validation was carried out separately on the validation datasets no. 11-20.

2.3 Automated segmentation of brain and tumor 2.3.1 General segmentation framework We adopted a general algorithm called adaptive template moderated classification (see [18, 19] and Appendix for details). The technique involves the iteration of statistical classification to assign labels to tissue types, and nonlinear registration to align (register) a digital anatomical atlas (a presegmented anatomical map) to the patient data (Figure 1). Statistical classification divides an image into different tissue classes based on the signal intensity value. If different tissue classes have the same or overlapping grey value distributions (e.g. CSF and fluid within the eyeballs), such methods fail. Therefore additional information about the spatial location of anatomical structures was derived from a registered anatomical atlas (manually segmented MRI of a single subject [6]). Objects of interest were identified on the classified images with local segmentation operations (mathematical morphology and region growing) [20].

2.3.2 Application to tumor segmentation For the task of brain tumor segmentation, the order in which the structures of interest were segmented followed a simple, hierarchical model of anatomy (Figure 2). By proceeding hierarchically from the outside to the inside of the head, each segmented structure defined a refined region of interest (ROI) for the next structure to be segmented. 5 different tissue classes were modeled: background, skin (fat/bone), brain, ventricles, and tumor. Due to the homogeneous tissue composition of meningiomas and low grade gliomas one tissue class was sufficient for the statistical model. An atlas of normal anatomy does not include pathologic structures. As a result, templates from the atlas were derived only for the head, the brain and the ventricles. First, the whole patient head was segmented from the background using thresholding and local segmentation strategies. Based on the segmentation of the head, an initial alignment of the atlas to the patient was established. Next, the intracranial cavity (ICC) was segmented from the head in two segmentation iterations (statistical classification, local segmentation strategy and re-registration of the atlas). At this point, all voxels belonging either to brain, ventricles and tumor were labeled as ICC. In the first iteration, the ICC was segmented using the head and the ICC template from the initially registered atlas. The atlas was then re-aligned based on the whole head and the ICC of the patient, followed by a second classification and local segmentation step. The ventricles were segmented from the ICC in a third segmentation iteration. At this point, the ICC only contained voxels belonging to the brain and the tumor. Having defined a region of interest for the tumor, which is located inside the brain and outside the ventricles and the skin (fat/bone), the tumor was segmented in two iteration cycles. In the first iteration, the tumor was classified using the anatomical knowledge from the atlas only, followed by application of the local segmentation strategy. Because there is no tumor template in the atlas, a straight-forward registration is not possible. Consequently, the tumor voxels were relabeled as ICC voxels prior to the registration 3 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

process. As a result, a spatial correspondence between the atlas and patient dataset could be established for every voxel, since the patient dataset contained no voxels labeled as tumor at the time the registration of the atlas was carried out. In the second iteration, the tumor segmentation from the first iteration was used as an anatomical template. Although this template was approximate, the additional information about the location of the tumor prevented from misclassifying voxels distant to the atlas template as tumor.

2.3.3 Initialization of the automated segmentation method To reduce the noise in the MRI without blurring object edges, an anisotropic diffusion filtering method was applied [21]. For the initialization of the automated segmentation method a graphical user interface was developed for the 2D display of MRI slices and the selection of example tissue points using a mouse (see Figure 3). The only interaction required by the operator (see section 2.5) was the selection of 3-4 example points for each tissue class, i.e. skin (fat/bone), brain, ventricles, and tumor. The program calculated a statistical model for the distribution of the gray values based on these manually selected tissue prototypes.

2.4 Manual segmentation of the brain and the brain tumor For manual segmentation of the brain and the tumor an interactive segmentation tool was used (MRX, GE Medical Systems, Schenectady, NY). The tool ran on an Ultra 10 workstation (Sun Microsystems, Mountain View, Calif). The structures were outlined slice-by-slice by human operators (see section 2.5) by pointing and clicking with a mouse. The program connected consecutive points with lines. An anatomical object was defined by a closed contour, and the program labeled every voxel of the enclosed volume.

2.5 Validation Experiments Due to the lack of an acceptable "gold standard" (e.g. a realistic phantom) to compare with, our definition of a segmentation "gold standard" was based upon the manual segmentations using interactive computer segmentation tools. However, manual segmentation is subject to interobserver variability and human error [6]. To minimize the influence of these factors whilst maintaining the means of measuring the segmentation accuracy of the individual raters, the standard was defined based on the segmentations of 4 independent human observers. A single 2D slice was randomly selected from the subset of the MRI volume that showed the tumor. On this slice, the brain and the tumor was then manually outlined by 4 human observers independently. The standard segmentation of brain and tumor in each patient dataset was defined as the area of those voxels where at least 3 out of 4 raters agreed on the identification. All other voxels were labeled as background. To assess the accuracy, the automated segmentation tool was trained once on a single MRI slice containing all tissue types of interest and executed on the full 3D dataset. This resulted in a segmentation of the entire dataset. For each dataset, the structures skin (fat/bone), brain, ventricles and tumor were segmented. The interrater variability of the 4 independent manual and the 4 independent automated segmentations was measured based on all 20 cases. For the measurement of the intraobserver variability, one of the medical experts also manually segmented the selected 2D slice 4 times over a period of one week in each of the 20 cases. Training of the automated method was also carried out 4 times over a period of one week for all 20 cases. During all experiments, the times for manual outlining, training and computation time of the automated

4 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

segmentation method was recorded.

2.6 Statistical Analysis The qualitative analysis was carried out on the basis of volume of overlap comparison with the "gold standard" (accuracy) and overall volume variability (reproducibility) in the 2D slice selected. Segmentation accuracy was defined as the percentage of correctly classified voxels (in object and background) with respect to the total number of voxels V in the image, i.e. (TP + TN) / V, where TP are the true positive and TN are the true negative voxels [22]. The mean and standard deviations (SD) of the accuracy values with respect to the 20 test cases were also calculated (Matlab version 4.1, Mathworks, Cambridge, MA). To assess the inter- and intrarater variability error, the coefficient of variation (CV% = 100 * [SD(volume) / MEAN(volume)] ) was calculated which does not measure correctness of segmentation but only the change in volume of objects in segmentations of different raters.

3 Results Examples of the manual and the automated segmentation of a meningioma (Figure 4) and a low grade glioma (Figure 5) indicate the similarity between the results with the two methods. The segmentation accuracy with the automated method was high and within the range of the accuracy of the manual method. The overall mean accuracy for the tumor segmentation over all 20 cases was 99.68 ± 0.29% (mean ± standard deviation) with the automated and 99.68 ± 0.24% with the manual method (Figure 6), while the mean accuracy for the brain segmentation over all 20 cases was 98.40 ± 0.57% and 98.81 ± 0.88% respectively (Figure 7). Intraobserver variability (Coefficients of variation, CV) for both the automated and the manual method was low. For brain and tumor segmentation, the mean intraobserver variability for all 20 cases with the automated method was 0.10-3.57% and 0.14-4.70%, while the manual method achieved CV values of 0.24-4.11% and 0.80-3.28% (Table 1, 2). Interobserver variability was lower with the automated than with the manual method. The mean interobserver variability for all 20 cases with the automated method was 0.33-4.72% and 0.99-6.11% for brain and tumor segmentation, while the manual method achieved CV values of 2.62-10.51% and 3.58-14.42% (Table 3, 4). The automated segmentation of a complete 3D image volume required approximately 75 minutes of unsupervised computation time (Sun ES 6000 server, 20 CPU's with 250 MHz, 5 GB of RAM; Sun Microsystems, Mountain View, CA). The overall operator time for training of the automated method was approximately 5-10 minutes (selection of example voxels for each of the relevant tissue classes). Manual outlining of brain and tumor required 1-3 minutes per slice. Manual segmentation of the 3D volume was in the order of 3-5 hours.

5 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

4 Discussion Our study shows that brain, meningiomas and low grade gliomas can be accurately and reproducibly segmented by means of automated processing of gradient echo MRI. We have shown that our algorithm allows complete segmentation of the brain and the tumor requiring only the manual selection of a small sample of example voxels (21-28). The goals of the development of automated segmentation tools are to make segmentation of MRI more practical by replacing manual outlining without a measurable effect on the results, reducing operator time, and to improve reproducibility. However, the validity of our segmentations is difficult to assess without the availability of a "gold standard". Therefore our validation study was designed to determine how closely the raters agreed within a single method (automated and manual) and how closely the segmentation results correlated using the two methods. The segmentation accuracy with the automated method was high, and within (maximum difference 0.6 %) the accuracy range with the manual method. The errors of the automated brain segmentation were in part due to over- and under-segmentation in the area of the tentorium and the lateral sulcus with abundant vessels. The algorithm tended to oversegment these areas if parts of the neck close to the cerebellum were misclassified as brain and the template ICC derived from the atlas was misaligned. The size of the structure affects the segmentation accuracy. Segmentation errors occur on the boundary of surfaces. Thus, the larger the surface of an object, the more voxels in the entire image can potentially be misclassified. Therefore, larger objects have a lower accuracy than smaller objects. Reproducibility was higher with the automated method, because only the selection of a few example points is required rather than making a decision on every voxel in the image during manual segmentation. The reproducibility of the brain and tumor segmentation was high. Nevertheless, the inter- and intraobserver reproducibility of both methods were higher for the brain than for the tumor. Larger objects tend to have a higher volumetric reproducibility than overall segmentation accuracy. Because the surface/volume ratio behaves approximately like 1/r the disagreement on voxel classes on the surface of larger objects with respect to the overall volume is less significant than for smaller objects. Interobserver variability was significantly reduced with the automated method. Manual interobserver variability is particularly high for low grade gliomas which are more difficult to segment causing deviating expert opinions. Automated segmentation is more robust to expert variation because it involves only the selection of typical example points for training of the algorithm while manual segmentation requires a human decision for every boundary voxel which is difficult due to e.g. the partial voluming. However, intraobserver variability was only improved for the meningioma segmentation. For low grade gliomas, manual intraobserver variability is significantly lower than interobserver variability because the execution of manual segmentation varies but not the opinion regarding the shape of the tumor. Therefore in comparison to the manual segmentation the automated method does not reduce interobserver variability significantly. Reproducibility is higher for meningiomas with both methods. This can be explained by comparing the gray value distributions of the meningiomas and the low grade gliomas respectively with the brain. The meningioma tissue class partially overlaps with parts of the skin, fat in the neck and the straight and superior sagittal sinus, and was well distinguishable from brain tissue with the application of contrast enhancing agent. Restricting the region of interest (ROI) to the ICC, the tissue that showed signal intensity

6 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

overlap with the meningioma was excluded and the meningioma could be successfully segmented. In some low grade glioma cases the ICC may not be a sufficient ROI for an accurate tumor segmentation due to similar signal intensities of the tumor and surrounding gray matter. False classifications cannot be corrected if brain misclassified as tumor tissue is adjacent to the tumor boundary (over-segmentation) or vice versa (under-segmentation). The incorporation of T2 weighted images, which clearly distinguish the tumor as hyper intense tissue, may enable the precise definition of the tumor boundaries. If the voxels of the brain misclassified as tumor are distant to the tumor boundary or connected to the tumor only by thin structures, or tumor voxels ``inside" the tumor are falsely classified as brain false classifications can be corrected. The algorithm developed in this work is based on template driven segmentation, where an anatomical atlas is used to guide a statistical classification process [24, 8, 9, 18, 19]. Clark et al. proposed a method for the automatic detection and segmentation of glioblastoma multiforme from a combination of T1, T2 and PD MRI using classification and an anatomical knowledge database, with accuracy greater than 90% [17]. Bonnie et al. recently reported results using an interactive tumor segmentation method [25]. However, the value is difficult to assess because no detail on the segmentation technique is given. Approaches based on MRI data alone using active contours [26] or multispectral classification [13, 14] work well if the tumor shows sufficient contrast to the brain. However, active contours require good initialization which is difficult to automate, while multispectral classifcation reveal problems with overlapping intensity distributions. The lack of automated segmentation methods requires tedious manual labor. This has been one of the reasons why 3D models have been typically limited to university research settings. The reduction of operator time (3-5 hours to 5-10 minutes) makes it practical to consider the integration of computerized segmentation into daily clinical routine as a service to neurosurgery. A technician carries out the initial work, and the result is verified by a radiologist while softreading the images. Our software currently runs on a large computer. However, these systems are becoming increasingly affordable [27]. In conclusion, we found that accurate segmentation is possible for meningiomas and low grade gliomas using our automated method. Further work is required to extend the tools to a broader range of brain tumors (e.g. glioblastoma multiforme). Future clinical studies on the accuracy and reproducibility of our technique based on a larger population will be necessary to determine the practical use for a clinical setting.

Appendix In the following we give the parameter settings and the features used. For algorithmic details see [18, 19]. The following parameter settings were used: Anisotropic diffusion filtering: 2 iterations, dt = 0.2, k = 5.2, kNN classification: k=5, number of classes C=5, affine registration: 9 degrees of freedom, 3 image resolution levels, distance transform: saturation distance = 100, non-linear registration: 3 resolution levels, window size w = 9×9×9, morphological operators: spherical element, w=7×7×7, region growing: 18-connectivity. 4 classification-registration iterations for the ICC segmentation were used, 1 iteration for the ventricle segmentation, and 2 iterations for the tumor segmentation. The brain and the ventricle are also re-segmented during tumor segmentation. For the segmentation of normal structures (i.e. skin/fat/bone, brain, ventricles), the pattern used in this work is vi = [v1i, ..., v5i]T, i being the index to a voxel location xi. The elements vji result from the image processing operations Tj, i.e.

7 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

v1i = T1(I(xi)), T1: Anisotropic Diffusion Filtering v2i = T2(A(xi)), T2: Distance Transform of skin/fat/bone v3i = T3(A(xi)), T3: Distance Transform of skin/fat/bone background v4i = T4(A(xi)), T4: Distance Transform of brain v5i = T5(A(xi)), T5: Distance Transform of brain background which are applied to the MRI image I(xi) and the image of the registered anatomical atlas A(xi) respectively. While T1 is carried out only during the preprocessing stage, the operators T2, ..., T5 are applied to the re-registered atlas in every segmentation iteration cycle. For the first tumor segmentation cycle, the patterns are also vi = [v1i, ..., v5i]T. For the second tumor segmentation cycle, the patterns are vi = [v1i, ..., v6i]T, where the vli, i=1..5 are defined as above, but with the additional pattern v6i = T6(A(xi)), T6: Distance transform of initial tumor segmentation where Ib is the result image of the first tumor segmentation.

Acknowledgments This work was supported (in part) by a grant from the Deutscher Akademischer Austauschdienst (DAAD). This investigation was supported (in part) by a Grant from the National Multiple Sclerosis Society (SW). This work was supported (in part) by NIH grants RO1 CA 46627-08, PO1 CA67165-01A1, PO1 AG04953-14, NSF grant BES 9631710 and Darpa grant F41624-96-2-0001. AN was supported by DFG grant NA 359/1-1. This work is part of the Dissertation of MK. The authors thank Rick Schwarz, MD, for helpful comments, and Fatma Ozlen, MD, for the help with the manual segmentations.

References [1] Jolesz FA. Image-guided procedures and the operating room of the future. Radiology 1997; 204:601-612. [2] Black PM, Moriarty T, Alexander E, et al. Development and implementation of intraoperative magnetic resonance imaging and its neurosurgical applications. Neurosurgery 1997; 41:831-845. [3] Nakajima S, Atsumi H, Bhalerao AH, et al. Computer-assisted surgical planning for cerebrovascular neurosurgery. Neurosurgery 1997; 41:403-409. [4] Hu X, Tan KK, Levin DN, et al. Three-dimensional magnetic resonance images of the brain: application to neurosurgical planning. J Neurosurg 1990; 72:433-440.

8 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

[5] Alexander E, Kikinis R, Jolesz FA. Intraoperative magnetic resonance imaging therapy. In: Barnett GH, Roberts D, Guthrie B, eds. Image-guided neurosurgery: Clinical applications of interactive surgical navigation. St. Louis, Quality Medical Publisher 1996; 260-266. [6] Kikinis R, Gleason PL, Moriarty TM, et al. Computer assisted interactive three-dimensional planning for neurosurgical procedures. Neurosurgery 1996; 38(4):640-651. [7] Cline HE, Lorensen E, Kikinis R, Jolesz F. Three-dimensional segmentation of MR images of the head using probability and connectivity. J Comput Assist Tomogr 1990; 14(6):1037-1045. [8] Vannier MW, Butterfield RL, Rickman DL, Jordan DM, Murphy WA, Biondetti PR. Multispectral magnetic resonance image analysis. Radiology 1985; 154:221-224. [9] Collins DL, Peters TM, Dai W, Evans AC. Model based segmentation of individual brain structures from MRI data. SPIE Visualization in Biomedical Computing 1992; 1808:10-23. [10] Kamber M, Shinghal R, Collins DL, et al. Model-based 3-D segmentation of multiple sclerosis lesions in magnetic resonance brain images. IEEE T Med Imaging 1995; 14(3):442-453. [11] Warfield SK, Dengler J, Zaers J, et al. Automatic identification of gray matter structures from MRI to improve the segmentation of white matter lesions. J Imag Guid Surg 1995; 1(6):326-338. [12] Gibbs P, Buckley DL, Blackband SJ, Horsman A. Tumor volume determination from MR images by morphological segmentation. Phys Med Biol 1996; 41:2437-2446. [13] Velthuizen RP, Clarke LP, Phuphanich S, et al. Unsupervised measurement of brain tumor volume on MR images. J Magn Reson Imaging 1995; 5:594-605. [14] Vinitski S, Gonzalez C, Mohamed F, et al. Improved intracranial lesion characterization by tissue segmentation based on a 3D feature map. Magn Reson Med 1997; 37:457-469. [15] Just M, Thelen M. Tissue characterization with T1, T2, and proton density values: Results in 160 patients with brain tumors. Radiology 1988; 169:779-785. [16] Just M, Higer HP, Schwarz M, et al. Tissue characterization of benign tumors: Use of NMR-tissue parameters. Magn Reson Imaging 1988; 6:463-472. [17] Clark M. Knowledge guided processing of magnetic resonance images of the brain. PhD thesis. University of South Florida 1998. [18] Warfield SK, Kaus MR, Jolesz FA, Kikinis R. Adaptive template moderated spatially varying statistical classification. In: Wells WH, Colchester A, Delp S., eds. Proceedings of the 1 st International Conference on Medical Image Computing and Computer-Assisted Intervention. Boston, MA 1998; 431-438. [19] Kaus MR, Warfield SK, Jolesz FA, Kikinis R. Segmentation of Meningiomas and Low Grade Gliomas in MRI. In: Taylor C, Colchester A, eds. Proceedings of the 2nd International Conference on Medical Image Computing and Computer-Assisted Intervention. Cambridge, England 1999; 1-10. [20] Serra J. Image analysis and mathematical morphology. London Academic, London, 1982.

9 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

[21] Gerig G, Kikinis R, Kübler O, et al. Nonlinear anisotropic filtering of MRI data. IEEE T Med Imaging 1992; 11(2):221-232. [22] Swets A, Pickett RM. Evaluation of Diagnostic Systems: Methods from Signal Detection Theory. Series in Cognition and Perception. Academic Press, 1982. [23] Clarke LP, Velthuizen RP, Phuphanich S, Schellenberg JD, Arrington JA, Silbinger M. MRI: Stability of three supervised segmentation techniques. Mag Reson Imaging 1993; 11(1):95-106. [24] Levin DN, Hu X, Tan KK, et al. The brain: Integrated three-dimensional display of MR and PET images. Radiology 1989; 172:783-789. [25] Bonnie NJ, Fukui MB, Meltzer CC, et al. Brain tumor volume measurement: Comparison of manual and semiautomated methods. Radiology 1999; (212):811-816. [26] Zhu H, Francis HY, Lam FK, Poon PWF. Deformable region model for locating the boundary of brain tumors. Proceedings of the IEEE 17th Annual Conference on Engineering in Medicine and Biology 1995; 411. [27] Kikinis R, Warfield SK, Westin CF. High performance computing (HPC) in medical image analysis (MIA) at the surgical planning laboratory (SPL). In: Proceedings of the 3rd High Performance Computing Asia Conference & Exhibition. Singapore, 1998.

Tables: Tumor Histology Meningioma Low Grade Glioma

Manual mean CV [%] SD 0.42 0.03 1.79 1.53

Automated mean CV [%] SD 0.36 0.45 1.44 1.33

Table 1: Intraobserver variability of brain volume segmented with the manual and the automated method averaged over all 20 cases. Tumor Histology Meningioma Low Grade Glioma

Manual mean CV [%] SD 1.58 0.98 2.08 0.78

Automated mean CV [%] SD 0.66 0.72 2.06 1.73

Table 2: Intraobserver variability of tumor volume segmented with the manual and the automated method averaged over all 20 cases .

10 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

Tumor Histology Meningioma Low Grade Glioma

file:////Tommy/bigweekly/marianna/kaus.html

Manual mean CV [%] SD 4.93 1.75 6.31 2.85

Automated mean CV [%] SD 1.84 0.65 2.71 1.68

Table 3: Interobserver variability of brain volume segmented with the manual and the automated method averaged over all 20 cases. Tumor Histology Meningioma Low Grade Glioma

Manual mean CV [%] SD 7.08 2.18 13.61 2.21

Automated mean CV [%] SD 2.66 0.38 2.97 1.58

Table 4: Interobserver variability of tumor volume segmented with the manual and the automated method averaged over all 20 cases.

Figures:

Figure 1: Tumor segmentation scheme

11 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

Figure 2: Flow diagram of the hierarchical segmentation proceeds from (a) to (d)

Figure 3: Graphical user interface for the automated segmentation method to allow the 2D display of MR slices and the selection of example tissue points using a mouse.

Figure 4: Example of an SPGR image with a meningioma (a), manual segmentation (b), statistical classification (c), and template moderated segmentation (d).

12 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

Figure 5: Example of manual and automated segmentation of a low grade gliomas: SPGR image (a), manual (b), and template moderated segmentation (c).

Figure 6: Brain segmentation accuracy of the manual (mean, minimum and maximum) and the automated method for each of the 20 brain tumor cases (Meningiomas: 1-3, 11, 12, 16; Low grade gliomas: 7-10, 13-15, 17-20). The accuracy with the automated method was consistent with the manual segmentation accuracy for most cases.

13 of 14

2/20/2001 12:24 PM

Automated Segmentation of MRI of Brain Tumors

file:////Tommy/bigweekly/marianna/kaus.html

Figure 7: Tumor segmentation accuracy of the manual (mean, minimum and maximum) and the automated method for each of the 20 brain tumor cases (Meningiomas: 1-3, 11, 12, 16; Low grade gliomas: 7-10, 13-15, 17-20). The accuracy with the automated method was consistent with the manual segmentation accuracy for most cases.

14 of 14

2/20/2001 12:24 PM