- Email: [email protected]
MRI Measurement of Brain Tumor Response: Comparison of Visual Metric and Automatic Segmentation
Magnetic Resonance Imaging, Vol. 16, No. 3, pp. 271–279, 1998 © 1998 Elsevier Science Inc. All rights reserved. Printed in the USA. 0730-725X/98 $19.00 1 .00
PII S0730-725X(97)00302-0
● Original Contribution
MRI MEASUREMENT OF BRAIN TUMOR RESPONSE: COMPARISON OF VISUAL METRIC AND AUTOMATIC SEGMENTATION LAURENCE P. CLARKE,* ROBERT P. VELTHUIZEN,* MATT CLARK,† JORGE GAVIRIA,* LARRY HALL,† DMITRY GOLDGOF,† REED MURTAGH,* S. PHUPHANICH,‡ AND STEVEN BREM‡ *Department of Radiology, College of Medicine, University of South Florida and the H. Lee Moffitt Cancer and Research Institute, Tampa, FL; †Department of Computer Science and Engineering, College of Engineering, University of South Florida, Tampa, FL; and the ‡Department of Neuro Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA An automatic magnetic resonance imaging (MRI) multispectral segmentation method and a visual metric are compared for their effectiveness to measure tumor response to therapy. Automatic response measurements are important for multicenter clinical trials. A visual metric such as the product of the largest diameter and the largest perpendicular diameter of the tumor is a standard approach, and is currently used in the Radiation Treatment Oncology Group (RTOG) and the Eastern Cooperative Oncology Group (EGOG) clinical trials. In the standard approach, the tumor response is based on the percentage change in the visual metric and is categorized into cure, partial response, stable disease, or progression. Both visual and automatic methods are applied to six brain tumor cases (gliomas) of varying levels of segmentation difficulty. The analyzed data were serial multispectral MR images, collected using MR contrast enhancement. A fully automatic knowledge guided method (KG) was applied to the MRI multispectral data, while the visual metric was taken from the MRI films using the T1 gadolinium enhanced image, with repeat measurements done by two radiologists and two residents. Tumor measurements from both visual and automatic methods are compared to ‘‘ground truth,’’ (GT) i.e., manually segmented tumor. The KG method was found to slightly overestimate tumor volume, but in a consistent manner, and the estimated tumor response compared very well to hand-drawn ground truth with a correlation coefficient of 0.96. In contrast, the visually estimated metric had a large variation between observers, particularly for difficult cases, where the tumor margins are not well delineated. The inter-observer variation for the measurement of the visual metric was only 16%, i.e., observers generally agreed on the lengths of the diameters. However, in 30% of the studied cases no consensus was found for the categorical tumor response measurement, indicating that the categories are very sensitive to variations in the diameter measurements. Moreover, the method failed to correctly identify the response in half of the cases. The data demonstrate that automatic 3D methods are clearly necessary for objective and clinically meaningful assessment of tumor volume in single or multicenter clinical trials. © 1998 Elsevier Science Inc. Keywords: MRI segmentation; Fuzzy clustering; Knowledge guided segmentation; Brain tumors; Therapy response.
INTRODUCTION
the tissues within the tumor bed to better delineate the active tumor margins as required for radiation treatment planning (RTP).1,4 – 6 In this work we are interested in the measurement of the relative changes in tumor volume during therapy, or more specifically, tumor response measurements.4 – 6 Tumor response measurement, however, poses a very difficult segmentation problem compared to all other applications because the segmentation
Magnetic resonance imaging (MRI) multispectral segmentation methods have been proposed for the determination of the volume of normal brain tissues (e.g., white or gray matter), multiple sclerosis (MS) lesions, and tumor volume.1–3 The advantages of multispectral methods, as opposed to single image gray scale methods such as seed growing, are the potential ability to differentiate RECEIVED 6/30/97; ACCEPTED 11/04/97. Address correspondence to Laurence P. Clarke, Ph.D., Professor of Radiology and Physics, Department of Radiology,
College of Medicine, 12901 Bruce B. Downs Blvd., MDC 17, Tampa, FL 33612-4799. E-mail: [email protected] 271
272
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
method may be dependent on: 1) tumor size and type; 2) stage, vascularity, and related degree of MR contrast enhancement; and 3) the varying influence of therapy on both normal and pathological tissues that may effect baseline MR relaxation parameters and the ability to serially differentiate tissues within the tumor bed.2 For example, the use of spatial knowledge as the primary means for segmentation (e.g., template matching or image warping), as proposed for normal tissue segmentations, is not suitable because there are no a priori spatial criteria that can be readily modeled for the changes in brain tumor images.7–9 The use of multispectral methods have therefore been proposed and specifically applied to each multispectral data set for each time interval; i.e., each of the patient image sets are used as their own control for application of the segmentation method. MRI segmentation methods to date for measuring tumor response have usually involved the use of supervised methods that require training regions of interest (ROIs) to be selected for each slice of the serial multispectral image data sets.2 The methods have included the use of Bayesian classifiers, neural networks or other pattern recognition methods.2,10 These methods have generally proved to be very dependent on the selection of the ROIs and are not optimum for tumor response measurements where reproducible serial segmentations are required. The poor performance of these supervised methods has often mistakenly been interpreted to be due to the RF non-uniformity.2,11 Although methods have been developed to reduce the dependence on the training ROIs such a semi supervised fuzzy clustering methods,10,12 where reduced inter- and intra-observer variability has been demonstrated for tumor response measures,4,6 these methods require significant operator input that is time consuming and are not suitable for multicenter trials. There is therefore a critical need for an fully automatic method for measuring tumor response. The use of visual metrics to estimate tumor response has been reported by several investigators13–16 and employed by various radiation oncology groups. The Radiation Treatment Oncology Group (RTOG) and Eastern Cooperative Oncology Group (ECOG)14 have used visual metrics to standardize response measurements and to reduce the inter-observer variability in tumor response rating. Response criteria rating definitions include 1) complete response, 2) partial response, 3) stable disease, or 4) progressive disease.14 Quantitative methods include maximum cross sectional area, perpendicular diameters of the cross sectional areas13–16 as evaluated in this work, or recently, volume measurements using seed growing methods.6 All of these approaches use the T1-enhanced image for estimating active tumor area and for estimating active tumor volume. Poor reproducibility of manual measurements, including insensitivity to response rating,
has been reported.14,17 Similarly, there is poor correlation between the visual response criteria and the above quantitative 2D and 3D metrics due to significant geometrical considerations and ill defined tumor margins when large case studies are performed.17 In many instances the visual response rating seriously underestimated changes in tumor volume, particularly for small lesions, multiple lesions, or poorly enhanced lesions that were not well visualized. These results strongly suggest the need for a fully automatic method for tumor response measures that specifically tracks changes, as opposed to an accurate or absolute measure of tumor volume as required for RTP planning.1 A comparison of these methods with automatic segmentation approaches for the same case study is a worthy goal. We therefore compare an automatic segmentation method against the use of a standard visual metric in use by ECOG trials, namely perpendicular diameters.14 The automatic method involves the use of fuzzy clustering methods (FCM) and a semi-supervised FCM (ssFCM), and incorporates very general knowledge from the N dimensional feature space domain and the anatomical or spatial domain to assist in the refinement of segmentation of tissues within the tumor bed, while still using the clustering methods as the primary segmentation method; this is referred to as the knowledge guided (KG) method.18 –21 This approach lends itself to the serial segmentation problem for measurement of tumor response, where the objective is to obtain consistent segmentation of tissues within the tumor bed as opposed to absolute determination of the tumor volume. A comparison of the KG method applied to the multispectral data and the visual metric from the MRI film hardcopy is performed for six case studies of glioma tumors of the brain of varying size and stage.5,6 A customized computer interface is used to determine ground truth (GT) for tumor volume to the proposed methods above, as has been previously reported.4 – 6,22,23 This work is organized as follows: Section 2 summarizes the KG method, the standard diameter based approach, how ‘‘ground truth’’ is established and the selection of the cases for clinical evaluation. An analysis of the variability of the diameter measurements, and a comparison of the results obtained with each of the methods are described in Section 3. The results are discussed and conclusions are presented in Section 4. METHODS 2.1 Clustering Methods for Segmentation The feasibility of using supervised segmentation methods such as ssFCM as applied to MRI MS serial data sets has been well demonstrated for reproducible measurement of relative changes in tumor volume for
MRI measurement of brain tumor response ● L.P. CLARKE
three different tumor types with varying degrees of staging and difficulty of segmentation.6 We have successfully demonstrated that FCM can provide good segmentation of white matter and gray matter in normal volunteers and of select tumor types and grades.24 However, as we progressed through a large image data base, we found that FCM did not perform well in difficult segmentation cases where tissues within the tumor bed were not well differentiated, i.e., results were tumor case dependent.5,19 In more recent work, a novel validity guided clustering (VGC) algorithm12,25 improved FCM partitions by using a split/merge criterion where tumor segmentation was improved for some difficult cases where FCM did not succeed. However, with further analysis of the extended data base it was clear that some form of knowledge guidance was necessary for robust performance to allow application in clinical trials.2 2.2 Overview of the KG Method Figure 1 shows schematically the different stages in the knowledge guided clustering approach.20,21 In each stage, clustering by fuzzy c-means is followed by an analysis using the rules in the knowledge base. The analysis allows the automatic labeling of clusters with their tissue type and the identification of the remaining clusters that need further processing. The approach, developed by Clark, Li, Goldgof and Hall,18 –21 is unique in that it couples sophisticated pattern recognition techniques that operate in the multispectral feature space with rules that encode knowledge about the relationships between tissue intensities and qualitative spatial characteristics of the MR images. In the literature several approaches have used knowledge in the spatial domain7–9 or image warping with the aid of a digital atlas26 –28 as the primary segmentation method with successful application to segmentation of normal brain tissues. However, these methods have not been described for the segmentation of brain tumors, particularly during therapy, because of the associated time related pathology changes and possible distortions from normal anatomy. The primary segmentation method in the KG approach is clustering of MRI data in the multi-dimensional feature space, while knowledge about the feature space and qualitative spatial knowledge of geometrical location or shape of tissues is used to fine tune the segmentation in a multi-stage refinement (Fig.1), an entirely different approach.20,21 2.3 Knowledge Guidance Criteria The stages in the knowledge guided approach in Fig. 1 consist of a fuzzy clustering phase followed by the application of a set of heuristic rules (knowledge). The rules are applied to the result of the fuzzy clustering to identify the tissue type of as many pixels as possible, based on the location of the cluster centers in feature space, and the
ET AL.
273
geometrical location and/or approximate shape of regions in the image. There are two types of rules. The first type uses the relative ordering of cluster centers in feature space.19,29 –31 This type of knowledge is really a set of relational features of FCM cluster centroids. There are also rules that create and manipulate histograms of the clusters. The second type of knowledge is spatial knowledge, which is used in the same way as the feature space knowledge to label classes obtained from clustering, but also to determine whether an adequate segmentation of the brain region was obtained. For example, some rules are based on an expected symmetry of the brain in the axial plane, which can be made operational in the knowledge guided system by comparing differences in the areas of normal tissues between the hemispheres. This is different from an atlas approach, where the MR data is compared to a standard image. In this case the information is extracted from the serial MRI data sets where the patient is used as its own control. 2.4 Patient Cases MRI data for six patients having glioma tumors (glioblastoma multiforme, grade III or higher grade astrocytoma) are investigated. The same cases were analyzed in other studies.6 Some patients previously received a combination of surgery and radiation therapy. During the 32-week monitoring period, each patient received chemotherapy, radiation therapy, or a combination of both. The patients were imaged on multiple occasions ranging from two to five scans depending on the patient case. 2.5 Imaging Protocol The transaxial multispectral MR images were acquired after administration of Gd-DTPA contrast enhancement using a 1.5 Tesla GE Signa Advantage MRI scanner with a multi-element head coil (General Electric Company, Milwaukee, WI). Contiguous 5 mm slice images were acquired with a field of view of either 240 mm (for Patients 1– 4) or 220 mm (for Patient 5) with a 256 3 192 acquisition matrix and were reconstructed to a 256 3 256 pixel image. The multispectral data set consisted of a 5 mm thick anatomical slice T1-weighted, proton-density-(PD) weighted, and a T2-weighted images. The T1 image was acquired using a standard spinecho (SE) sequence with a repetition time (TR)/echo time (TE) 5 650/11 ms. The PD and T2 images were acquired using a fast spin-echo (FSE) sequence with a TR/TEeff 5 4000/17 ms for PD image and a TR/TEeff 5 4000/102 ms for T2 image. All images were processed on a Sun workstation, typically on one of the Sun SPARC 10 or 20 computer systems or the Sun 4/630 server (Sun Microsystems, Inc., Mountain View, California), available in our digital medical imaging laboratory.
274
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
Fig. 1. Overview of the stages in the knowledge-guided segmentation. Each stage performs first a fuzzy clustering; the clusters are then analyzed by the knowledgebase in feature space and on spatial characteristics. Each stage identifies one or more tissue classes and generates a mask for continued processing.
2.6 Diameter Measurements The Eastern Cooperative Oncology Group is one of the national cooperative groups that established carefully
designed response criteria.14 Brain tumor response is measured on computer tomography or MRI by measuring the greatest diameter of the tumor cross section and
MRI measurement of brain tumor response ● L.P. CLARKE
ET AL.
275
physicians then measured diameters on the preselected image slice.
Fig. 2. Choosing the largest diameter is ambiguous, and variability between operators will exist.
its largest perpendicular diameter. The area of the lesion(s) is obtained by multiplying the greatest diameter by its largest perpendicular diameter. This area (or sum of areas if there are multiple lesions) is then used to classify the response into one of four classes: 1) a complete response (CR) is defined as the disappearance of all evidence of active tumor; 2) a partial response (PR) is defined as at least a 50% decrease in the product of the cross sectional diameters; 3) patients with less than a 50% decrease or a 25% increase in the product of the diameters are considered to have stable disease (SD); and 4) progressive disease (PD) is defined as an increase of more than 25% in the product of the cross sectional diameters, or the appearance of new lesions. Three physicians independently measured the greatest diameter and the largest perpendicular diameter on the cross sectional images. Each was presented with all the films with the MR images on a film formatter. The physicians were asked to evaluate each MR data set by choosing the image slice where the lesion showed the largest diameter, and to use that image to do the measurements. The diameters were measured using a compass and the centimeter-scale printed on the film. Considerable variability should be expected with the method. For example, in Fig. 2 a slice is shown where the largest diameter of a single lesion is present. However, physicians may not make the same choice when deciding the location of the tumor boundary. The variability using this method may be partly caused by the choice of image slice. To evaluate the uncertainty due to the interpretation of a single image, we selected from each patient scan one T1-weighted image slice showing the lesion. Four
2.7 GT Criteria Validation of segmentation results for tumor response assessments requires establishing ‘‘ground truth’’ with respect to the tumor volume changes as can be seen on the MR images. Ideally a surgical procedure would provide the true volume, but a single measurement in time would not provide a means for verification of relative changes in tumor volume.2 GT was therefore established using a custom design interface to display full multispectral MRI images and a transparent overlay of the physician determined GT to allow reproducible hand drawing of tumor tissue for each 2D slice. Physician experts (neuroradiologists) generate GT on each slice through the tumor volumes, a very time consuming task. The variation using this method was proven to be small at less than 5%,23 with the source of variation being uncertainty in the image rather than labeling precision. This method has been successfully used in our laboratory for the last 3 years4 – 6,22 for verifying GT specifically for tumor response measurements. However, the method is still based primarily on visual estimates from the T1weighted contrast enhanced image and may not fully represent absolute tumor size or margins. This method should be distinguished from visual extracted 2D methods described above, because this GT method, although manual, requires precision on the voxel level. 2.8 Comparison of the Methods In the experiments we obtained several measurements as is illustrated in Fig. 3. First, physicians measured diameters using a compass on the films of the studies cases. These measurements were used to calculate the product of the diameters, effectively obtaining an ‘‘area’’ enclosing the tumor. Several physicians conducted the measurements, so that an inter-observer variability for both individual diameter measurements and the area measurements could be calculated. As described in section 2.6, the area measurement of a post-treatment MRI study is then compared to the area measurement of the baseline MRI study, and a response class is assigned based on the percentage change of the area. In this study, we also evaluated the inter-observer variability for the response class (Section 3.4). The knowledge guided system and the manually defined GT both yield tumor volumes in cm3. A comparison of the physician determined ‘‘area’’ measurement (in cm2) can only be done in terms of the ranking of the measurements because they have a different unit. To compare response classes, we define decision boundaries on the change in tumor volume. In previous experiments we found that the variability of segmenting MRI data is
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
276
Fig. 3. Measurements obtained for each MRI data set in the study. The visual assessment is done by several physicians, each measuring the largest diameter and the largest perpendicular diameter. The product of these diameters represent an ‘‘area’’ which forms the basis for the response class on follow-up studies. The knowledge guided system automatically segments the whole tumor, yielding a volume. The ground truth method involves manual segmentation of each pixel, also yielding a volume. These volumes are then used for a response class assignment for follow-up studies.
about 5%.4 Therefore, we propose detection of partial response when the tumor shrinks by more than 10%, or progressive disease when the tumor grows by more than 10%. RESULTS 3.1 Area Measurements Three physicians independently measured the greatest diameter and the largest perpendicular diameter on the original MRI films. The product of the measured diameters for each physician is obtained as shown in Fig. 4. The variability is immediately apparent. The coefficient of variation (CoV) is defined as the percent standard # deviation sX of a set of measurements X of the mean X of X:
periment where the physicians picked the image themselves (p 5 0.44). 3.2 Diameter Measurements Because the second set of measurements were done on corresponding images, the measured diameters can be compared. We found that the average CoV for the largest diameter was 10%, where the average CoV for the largest perpendicular diameter was 12%, where the difference is non-significant (p 5 0.30). For both diameters the
sx CoV(X) 5 100% 3 # X For each of the 17 volumes studied, the CoV over the physician measurements was calculated. The average CoV for area measurements was 16%. To gain insight in the source of this variability, a second experiment was done, where four physicians were asked to estimate the diameters on a pre-selected slice from each patient MRI data set. This eliminates the variability due to the choice of slice. The average CoV over the 17 data sets for the area was 19%, where there was no significant difference with the first ex-
Fig. 4. Area measurements for five patients by three physicians. Each patient case has at least the baseline and one follow-up MRI data set evaluated.
MRI measurement of brain tumor response ● L.P. CLARKE
ET AL.
277
Table 1. Spearman correlation coefficients RS for area measurements with ground truth volumes (p 5 significance)
RS p
Fig. 5. Relationship between the manually determined tumor volume and the area measurements done by three physicians.
average variability was 60.5 cm (corresponding to about 1.6 mm on the film). The average correlation coefficient for the diameter measurements was 0.26 6 0.38, indicating that the diameter measurements are independent, i.e., if a physician overestimated the largest diameter (compared to the average), the other diameter was not necessarily also overestimated. In other words, the estimation of the location of the tumor boundary (e.g., see Fig. 2) appears to be the major source of variability, rather then the choice of image or the characteristics of the enhancement, as the latter would be reflected in larger correlation coefficients. 3.3 Comparison with GT Figure 5 depicts the relationship between the GT volume and the physician area measurements, based on the data in the first experiment where each physician selected the image slice to do the measurements on. There is a substantial variability, but more importantly, the relationship with GT is weak. It is unlikely that the relationship can be modeled as a straight line, because the area measurements have a different dimension than the volumes do. The appropriate test statistic is then the Spearman rank-order correlation coefficient RS, a nonparametric test describing the relation: ‘‘if the GT is larger, then the area measurement is larger as well.’’ The results are listed in Table 1. The aggregate for all physician measurements is RS 5 0.39 (p 5 0.005), which indicates that the area measurements, even taken by multiple physicians, do not represent the tumor volume well. For comparison, the Spearman correlation coefficient for GT with the automatically obtained tumor volume using the hybrid knowledge guided segmentation method is RS 5 0.90 (p , 1025). Because GT data and knowledge guided results both
JG
RM
EH
0.49 0.04
0.28 0.28
0.48 0.05
express the tumor volume, the more powerful correlation coefficient for a straight line through the origin can be calculated: RO 5 0.96, showing that the KG system allows accurate response measurements. This is also illustrated in Fig. 6, where the range of physician measurements is shown in gray, and the GT and knowledge guided volumes are drawn in with a separate axis on the right hand side. Where Fig. 5 represents the overall relationship between tumor volume and measured area, Fig. 6 shows that relationship on a patient-by-patient basis. 3.4 Response Measurements In the tumor response measurement protocol, the area measurements are used to establish a response class. Although there is little correlation between area measurements and GT volume, the area measurements may still be adequate for estimation of the response class. Comparison with the volumes found through manual labeling or the knowledge based approach is not straightforward, as the measures represent a different dimension. For purposes of evaluation, we classified a volume measurement as Partial Response when the tumor volume
Fig. 6. Comparisons of response patterns for physician determined areas, knowledge based segmentation and manually determined ground truth volume. The scale for area measurements is on the left y-axis, for volumes the scale is on the right y-axis.
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
278
Table 2. Response measurements for each follow up (FU) Patient Patient FU2 FU3 FU4 Patient Patient Patient FU2 Patient FU2 FU3 FU4
1, FU1
2, FU1 3, FU1 4, FU1 5, FU1
GT
KG
JG
RM
EH
SD SD PR SD PD PR SD PD PR PR SD PR
SD SD SD SD PD PR SD PD PR PR SD PR
SD SD SD SD SD SD SD PD SD SD SD SD
SD SD SD SD SD SD SD PD PR PR SD PR
SD SD SD SD PD SD SD PD SD SD SD SD
GT 5 ground truth; KG 5 knowledge guided derived volume; JG,RM and EH: response based on physician area measurements; SD 5 stable disease; PR 5 partial response; PD 5 progressive disease.
reduced by more than 10% compared to the baseline image, and we classified the volume measurement as Progressive Disease when the tumor volume increased by more than 10% compared to the baseline. These percentages are smaller than those used for the area measurements, but are reasonable because we found variabilities in tissue volumes on volunteers, when they were imaged serially over similar periods as the brain tumor patients were imaged, to be less than 5%.4 A 10% threshold for the classification is therefore a prudent but warranted decision boundary. As can be seen in Table 2, the three physicians agreed in eight cases on the response, seven of which were ‘‘stable disease’’ according to the criteria for area measurements. However, in only six of those eight times, the area based classification corresponded to the GT volumebased class. In contrast, the knowledge guided segmentation agreed on 11 cases with the GT volume, where for the one case of disagreement the difference in tumor response was only 4%. DISCUSSION AND CONCLUSIONS In this work, we have shown that the protocol for measuring tumor response as commonly used in clinical trials for various cooperative groups does not allow accurate or even correct estimation of the tumor response as can be measured using the gadolinium enhancement in T1-weighted images. The variability in the measurement of the diameters is not the main source of error, because it is only about 10%, which compounds into variability in area measurements of 16%. It is the hypothesis that is the basis for the protocol—that the largest area represents tumor size—that we found to be untrue. At the same time, we have shown that a previously reported auto-
matic segmentation,18 –21 which does not require any operator or physician input, strongly correlates with the tumor volume and, hence, is a viable candidate to replace the area measurements. Acknowledgments–The authors thank Drs. E. Harris, D. Stallworth and G. Fernandez for their help in the film reading and diameter measurements. This work was supported in part by a grant from the National Cancer Institute (R01-CA5942501).
REFERENCES 1. Chaney, E.L.; Pizer, S.M. Defining anatomical structures from medical images. Semin. Radiat. Oncol. 2:215–225; 1992. 2. Clarke, L.P.; Velthuizen, R.P.; Camacho, M.; Heine, J.J.; Vaydianathan, M.; Hall, L.O.; Thatcher, R.W.; Silbiger, M.L. Review of M.A., MRI segmentation: Methods and application. Magn. Reson. Imaging 12:342–368; 1995. 3. Johnston, B.; Atkins, M.S.; Mackiewich, B.; Anderson, M. Segmentation of multiple sclerosis lesions in intensity corrected multispectral MRI. IEEE Trans. Med. Imaging 15: 154 –169; 1996. 4. Vaidyanathan, M.; Heidtman, C.; Clarke, L.P.; Hall, L.O., et al. Normal brain tissue volume measurements using multispectral MRI segmentation methods. Magn. Reson. Imaging 15:87–97; 1997. 5. Velthuizen, R.P.; Clarke, L.P.; Phuphanich, S.; Hall, L.O.; Bensaid, A.M.; Arrington, J.A.; Greenberg, H.M.; Silbiger, M.L. Unsupervised measurement of brain tumor volume on MR images. J. Magn. Reson. Imaging 5:594 – 605; 1995. 6. Vaidyanathan, M.; Clarke, L.P.; Hall, L.O.; Heidtman, C.; Velthuizen, R.P.; Gosche, K.; Phuphanich, S.; Wagner, H.; Greenburg, H.; Silbiger, M.L. Monitoring brain tumor response to therapy using MRI segmentation. Magn. Reson. Imaging 3:323–334; 1997. 7. Sonka, M.; Tadikonda, S.K.; Collins, S.M. Knowledgebased interpretation of MR brain images. IEEE Trans. Med. Imaging 15:443– 452; 1996. 8. Li, X.; Bhide, S.; Kabuka, M.R. Labeling of MR brain images using boolean neural network. IEEE Trans. Med. Imaging 15:628 – 638; 1996. 9. Amit, Y. Graphical shape templates for automatic anatomy detection with applications to MRI brain scans. IEEE Trans. Med. Imaging 16:28 – 40; 1997. 10. Bensaid, A.M.; Hall, L.O. et al. Partially supervised clustering for image segmentation. Pattern Recognition 29: 859 – 871; 1996. 11. Velthuizen, R.P.; Heine, J.J.; Clarke, L.P.; Lin, H. and Fletcher, L.M. Review and Evaluation of MRI nonuniformity corrections for brain tumor response measurements. Medical Physics. In press, 1998. 12. Bensaid, A.M. Improved fuzzy clustering for pattern recognition with applications to image segmentation. Ph.D. Dissertation. Tampa, FL: University of South Florida; 1994. 13. Zijlstra, E.J.; Taphoorn, M.J.B.; Barkhof, F.; Hoogenraad, F.G.C.; Heimans, J.J.; Valk, J. Radiotherapy response of
MRI measurement of brain tumor response ● L.P. CLARKE
14.
15.
16.
17. 18.
19.
20.
21.
22.
cerebral metastasis quantified by serial MR imaging. J. Neuro-Oncology 21:171–176; 1994. Grossman, S.A.; Burch, P.A. Quantitation of tumor response to anti-neoplastic therapy. Semin. Oncol. 15:441– 454; 1988. Watling, C.J.; Lee, D.H.; Macdonald, D.R.; Cairncross, J.G. Corticosteroid-induced magnetic resonance imaging changes in patients with recurrent malignant glioma. J. Clin Oncol. 12:1886 –1889; 1994. Buckner, J.C.; Brown, L.D.; Kugler, J.W.; Cascino, T.L.; Krook, J.E.; Mailliard, J.A.; Kardinal, C.G.; Tschetter, L.K.; O’Fallon, J.R.; Scheithauer, B.W. Phase II evaluation of recombinant interferon alpha and bcnu in recurrent glioma. J. Neurosurg. 82:430 – 435; 1995. Laperrire, N.J.; Berstein, M. Radiotherapy for brain tumors. Cancer J. Clin. 4:96 –108; 1994. Li, C.; Goldgof, D.B.; Hall, L.O. Automatic segmentation and tissue labeling of MR brain images. IEEE Trans. Med. Imaging 12:740 –750; December 1993. Clark, M.C.; Hall, L.O.; Goldgof, D.B.; Clarke, L.P.; Velthuizen, R.P.; Silbiger, M.S. MRI segmentation using fuzzy clustering techniques. IEEE Engineering in Medicine and Biology Magazine 13:730 –742; 1994. Clark, M.C.; Hall, L.O.; Goldgof, D.B. Using fuzzy information in knowledge guided segmentation of brain tumors. In: Fuzzy Logic in AI Workshop at IJCAI, Montreal, CA; August 1995. Clark, M.C.; Hall, L.O.; Goldgof, D.B.; Velthuizen, R.P.; Silbiger, M.S. Automatic tumor segmentation using knowledge-based techniques. IEEE Trans. Med. Imaging In press, 1998. Vaidyanathan, M.; Clarke, L.P.; Velthuizen, R.P.; Phuphanich, S.; Bensaid, A.M.; Hall, L.O.; Bezdek, J.C.; Silbiger, M. Comparison of supervised MRI segmentation
23.
24.
25.
26.
27.
28.
29.
30.
31.
ET AL.
279
methods for tumor volume determination during therapy. Magn. Reson. Imaging 13:719; 1995. Velthuizen, R.P.; Clarke, L.P. An interface for validation of MR image segmentation. In: Proceedings of the 16th Annual International Conference of the IEEE Engineering Medicine and Biology Society; 1994:547–548. Hall, L.O.; Bensaid, A.; Clarke, L.P.; Velthuizen, R.P.; Silbiger, M.L.; Bezdek, J.C. A comparison of neural network and fuzzy clustering techniques in segmenting magnetic resonance images of the brain. IEEE Trans. Neural Networks 3:672– 682; 1992. Bensaid, A.M.; Hall, L.O.; Bezdek, J.C.; Clarke, L.P. Validity-guided (re)clustering for image segmentation. IEEE Trans. Fuzzy Systems 4:112–123; 1996. Kamber, M.; Shingal, R.; Collins, D.L.; Francis, G.S.; Evans, A.C. Model-based 3D segmentation of multiple sclerosis lesions in magnetic resonance brain images. IEEE Trans. Med. Imaging 14:442– 453; 1995. Thompson, P.; Toga, A.W. A surface based technique for warping three-dimensional images of the brain. IEEE Trans. Med. Imaging 15:402– 417; 1996. Sandor, S.; Leahy, R. Surface-based labeling of cortical anatomy using a deformable atlas. IEEE Trans. Med. Imaging 16:41–54; 1997. Galloway, R.L.; Maciunas, R.J.; Failinger, A.L. Factors affecting perceived tumor volumes in magnetic resonance imaging. Ann. Biomed. Engineering 21:367–375; 1993. Bronen, R.; Sze, G. Magnetic resonance imaging contrast agents: Theory and application to the central nervous system. J. Neurosurgy 73:820 – 839; 1990. Bradley, W.G.; Yuh, W.T.C.; Bydder, G.M. Use of MR imaging contrast agents in the brain. J. Magn. Reson. Imaging 3:199 –218; 1993.
PII S0730-725X(97)00302-0
● Original Contribution
MRI MEASUREMENT OF BRAIN TUMOR RESPONSE: COMPARISON OF VISUAL METRIC AND AUTOMATIC SEGMENTATION LAURENCE P. CLARKE,* ROBERT P. VELTHUIZEN,* MATT CLARK,† JORGE GAVIRIA,* LARRY HALL,† DMITRY GOLDGOF,† REED MURTAGH,* S. PHUPHANICH,‡ AND STEVEN BREM‡ *Department of Radiology, College of Medicine, University of South Florida and the H. Lee Moffitt Cancer and Research Institute, Tampa, FL; †Department of Computer Science and Engineering, College of Engineering, University of South Florida, Tampa, FL; and the ‡Department of Neuro Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA An automatic magnetic resonance imaging (MRI) multispectral segmentation method and a visual metric are compared for their effectiveness to measure tumor response to therapy. Automatic response measurements are important for multicenter clinical trials. A visual metric such as the product of the largest diameter and the largest perpendicular diameter of the tumor is a standard approach, and is currently used in the Radiation Treatment Oncology Group (RTOG) and the Eastern Cooperative Oncology Group (EGOG) clinical trials. In the standard approach, the tumor response is based on the percentage change in the visual metric and is categorized into cure, partial response, stable disease, or progression. Both visual and automatic methods are applied to six brain tumor cases (gliomas) of varying levels of segmentation difficulty. The analyzed data were serial multispectral MR images, collected using MR contrast enhancement. A fully automatic knowledge guided method (KG) was applied to the MRI multispectral data, while the visual metric was taken from the MRI films using the T1 gadolinium enhanced image, with repeat measurements done by two radiologists and two residents. Tumor measurements from both visual and automatic methods are compared to ‘‘ground truth,’’ (GT) i.e., manually segmented tumor. The KG method was found to slightly overestimate tumor volume, but in a consistent manner, and the estimated tumor response compared very well to hand-drawn ground truth with a correlation coefficient of 0.96. In contrast, the visually estimated metric had a large variation between observers, particularly for difficult cases, where the tumor margins are not well delineated. The inter-observer variation for the measurement of the visual metric was only 16%, i.e., observers generally agreed on the lengths of the diameters. However, in 30% of the studied cases no consensus was found for the categorical tumor response measurement, indicating that the categories are very sensitive to variations in the diameter measurements. Moreover, the method failed to correctly identify the response in half of the cases. The data demonstrate that automatic 3D methods are clearly necessary for objective and clinically meaningful assessment of tumor volume in single or multicenter clinical trials. © 1998 Elsevier Science Inc. Keywords: MRI segmentation; Fuzzy clustering; Knowledge guided segmentation; Brain tumors; Therapy response.
INTRODUCTION
the tissues within the tumor bed to better delineate the active tumor margins as required for radiation treatment planning (RTP).1,4 – 6 In this work we are interested in the measurement of the relative changes in tumor volume during therapy, or more specifically, tumor response measurements.4 – 6 Tumor response measurement, however, poses a very difficult segmentation problem compared to all other applications because the segmentation
Magnetic resonance imaging (MRI) multispectral segmentation methods have been proposed for the determination of the volume of normal brain tissues (e.g., white or gray matter), multiple sclerosis (MS) lesions, and tumor volume.1–3 The advantages of multispectral methods, as opposed to single image gray scale methods such as seed growing, are the potential ability to differentiate RECEIVED 6/30/97; ACCEPTED 11/04/97. Address correspondence to Laurence P. Clarke, Ph.D., Professor of Radiology and Physics, Department of Radiology,
College of Medicine, 12901 Bruce B. Downs Blvd., MDC 17, Tampa, FL 33612-4799. E-mail: [email protected] 271
272
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
method may be dependent on: 1) tumor size and type; 2) stage, vascularity, and related degree of MR contrast enhancement; and 3) the varying influence of therapy on both normal and pathological tissues that may effect baseline MR relaxation parameters and the ability to serially differentiate tissues within the tumor bed.2 For example, the use of spatial knowledge as the primary means for segmentation (e.g., template matching or image warping), as proposed for normal tissue segmentations, is not suitable because there are no a priori spatial criteria that can be readily modeled for the changes in brain tumor images.7–9 The use of multispectral methods have therefore been proposed and specifically applied to each multispectral data set for each time interval; i.e., each of the patient image sets are used as their own control for application of the segmentation method. MRI segmentation methods to date for measuring tumor response have usually involved the use of supervised methods that require training regions of interest (ROIs) to be selected for each slice of the serial multispectral image data sets.2 The methods have included the use of Bayesian classifiers, neural networks or other pattern recognition methods.2,10 These methods have generally proved to be very dependent on the selection of the ROIs and are not optimum for tumor response measurements where reproducible serial segmentations are required. The poor performance of these supervised methods has often mistakenly been interpreted to be due to the RF non-uniformity.2,11 Although methods have been developed to reduce the dependence on the training ROIs such a semi supervised fuzzy clustering methods,10,12 where reduced inter- and intra-observer variability has been demonstrated for tumor response measures,4,6 these methods require significant operator input that is time consuming and are not suitable for multicenter trials. There is therefore a critical need for an fully automatic method for measuring tumor response. The use of visual metrics to estimate tumor response has been reported by several investigators13–16 and employed by various radiation oncology groups. The Radiation Treatment Oncology Group (RTOG) and Eastern Cooperative Oncology Group (ECOG)14 have used visual metrics to standardize response measurements and to reduce the inter-observer variability in tumor response rating. Response criteria rating definitions include 1) complete response, 2) partial response, 3) stable disease, or 4) progressive disease.14 Quantitative methods include maximum cross sectional area, perpendicular diameters of the cross sectional areas13–16 as evaluated in this work, or recently, volume measurements using seed growing methods.6 All of these approaches use the T1-enhanced image for estimating active tumor area and for estimating active tumor volume. Poor reproducibility of manual measurements, including insensitivity to response rating,
has been reported.14,17 Similarly, there is poor correlation between the visual response criteria and the above quantitative 2D and 3D metrics due to significant geometrical considerations and ill defined tumor margins when large case studies are performed.17 In many instances the visual response rating seriously underestimated changes in tumor volume, particularly for small lesions, multiple lesions, or poorly enhanced lesions that were not well visualized. These results strongly suggest the need for a fully automatic method for tumor response measures that specifically tracks changes, as opposed to an accurate or absolute measure of tumor volume as required for RTP planning.1 A comparison of these methods with automatic segmentation approaches for the same case study is a worthy goal. We therefore compare an automatic segmentation method against the use of a standard visual metric in use by ECOG trials, namely perpendicular diameters.14 The automatic method involves the use of fuzzy clustering methods (FCM) and a semi-supervised FCM (ssFCM), and incorporates very general knowledge from the N dimensional feature space domain and the anatomical or spatial domain to assist in the refinement of segmentation of tissues within the tumor bed, while still using the clustering methods as the primary segmentation method; this is referred to as the knowledge guided (KG) method.18 –21 This approach lends itself to the serial segmentation problem for measurement of tumor response, where the objective is to obtain consistent segmentation of tissues within the tumor bed as opposed to absolute determination of the tumor volume. A comparison of the KG method applied to the multispectral data and the visual metric from the MRI film hardcopy is performed for six case studies of glioma tumors of the brain of varying size and stage.5,6 A customized computer interface is used to determine ground truth (GT) for tumor volume to the proposed methods above, as has been previously reported.4 – 6,22,23 This work is organized as follows: Section 2 summarizes the KG method, the standard diameter based approach, how ‘‘ground truth’’ is established and the selection of the cases for clinical evaluation. An analysis of the variability of the diameter measurements, and a comparison of the results obtained with each of the methods are described in Section 3. The results are discussed and conclusions are presented in Section 4. METHODS 2.1 Clustering Methods for Segmentation The feasibility of using supervised segmentation methods such as ssFCM as applied to MRI MS serial data sets has been well demonstrated for reproducible measurement of relative changes in tumor volume for
MRI measurement of brain tumor response ● L.P. CLARKE
three different tumor types with varying degrees of staging and difficulty of segmentation.6 We have successfully demonstrated that FCM can provide good segmentation of white matter and gray matter in normal volunteers and of select tumor types and grades.24 However, as we progressed through a large image data base, we found that FCM did not perform well in difficult segmentation cases where tissues within the tumor bed were not well differentiated, i.e., results were tumor case dependent.5,19 In more recent work, a novel validity guided clustering (VGC) algorithm12,25 improved FCM partitions by using a split/merge criterion where tumor segmentation was improved for some difficult cases where FCM did not succeed. However, with further analysis of the extended data base it was clear that some form of knowledge guidance was necessary for robust performance to allow application in clinical trials.2 2.2 Overview of the KG Method Figure 1 shows schematically the different stages in the knowledge guided clustering approach.20,21 In each stage, clustering by fuzzy c-means is followed by an analysis using the rules in the knowledge base. The analysis allows the automatic labeling of clusters with their tissue type and the identification of the remaining clusters that need further processing. The approach, developed by Clark, Li, Goldgof and Hall,18 –21 is unique in that it couples sophisticated pattern recognition techniques that operate in the multispectral feature space with rules that encode knowledge about the relationships between tissue intensities and qualitative spatial characteristics of the MR images. In the literature several approaches have used knowledge in the spatial domain7–9 or image warping with the aid of a digital atlas26 –28 as the primary segmentation method with successful application to segmentation of normal brain tissues. However, these methods have not been described for the segmentation of brain tumors, particularly during therapy, because of the associated time related pathology changes and possible distortions from normal anatomy. The primary segmentation method in the KG approach is clustering of MRI data in the multi-dimensional feature space, while knowledge about the feature space and qualitative spatial knowledge of geometrical location or shape of tissues is used to fine tune the segmentation in a multi-stage refinement (Fig.1), an entirely different approach.20,21 2.3 Knowledge Guidance Criteria The stages in the knowledge guided approach in Fig. 1 consist of a fuzzy clustering phase followed by the application of a set of heuristic rules (knowledge). The rules are applied to the result of the fuzzy clustering to identify the tissue type of as many pixels as possible, based on the location of the cluster centers in feature space, and the
ET AL.
273
geometrical location and/or approximate shape of regions in the image. There are two types of rules. The first type uses the relative ordering of cluster centers in feature space.19,29 –31 This type of knowledge is really a set of relational features of FCM cluster centroids. There are also rules that create and manipulate histograms of the clusters. The second type of knowledge is spatial knowledge, which is used in the same way as the feature space knowledge to label classes obtained from clustering, but also to determine whether an adequate segmentation of the brain region was obtained. For example, some rules are based on an expected symmetry of the brain in the axial plane, which can be made operational in the knowledge guided system by comparing differences in the areas of normal tissues between the hemispheres. This is different from an atlas approach, where the MR data is compared to a standard image. In this case the information is extracted from the serial MRI data sets where the patient is used as its own control. 2.4 Patient Cases MRI data for six patients having glioma tumors (glioblastoma multiforme, grade III or higher grade astrocytoma) are investigated. The same cases were analyzed in other studies.6 Some patients previously received a combination of surgery and radiation therapy. During the 32-week monitoring period, each patient received chemotherapy, radiation therapy, or a combination of both. The patients were imaged on multiple occasions ranging from two to five scans depending on the patient case. 2.5 Imaging Protocol The transaxial multispectral MR images were acquired after administration of Gd-DTPA contrast enhancement using a 1.5 Tesla GE Signa Advantage MRI scanner with a multi-element head coil (General Electric Company, Milwaukee, WI). Contiguous 5 mm slice images were acquired with a field of view of either 240 mm (for Patients 1– 4) or 220 mm (for Patient 5) with a 256 3 192 acquisition matrix and were reconstructed to a 256 3 256 pixel image. The multispectral data set consisted of a 5 mm thick anatomical slice T1-weighted, proton-density-(PD) weighted, and a T2-weighted images. The T1 image was acquired using a standard spinecho (SE) sequence with a repetition time (TR)/echo time (TE) 5 650/11 ms. The PD and T2 images were acquired using a fast spin-echo (FSE) sequence with a TR/TEeff 5 4000/17 ms for PD image and a TR/TEeff 5 4000/102 ms for T2 image. All images were processed on a Sun workstation, typically on one of the Sun SPARC 10 or 20 computer systems or the Sun 4/630 server (Sun Microsystems, Inc., Mountain View, California), available in our digital medical imaging laboratory.
274
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
Fig. 1. Overview of the stages in the knowledge-guided segmentation. Each stage performs first a fuzzy clustering; the clusters are then analyzed by the knowledgebase in feature space and on spatial characteristics. Each stage identifies one or more tissue classes and generates a mask for continued processing.
2.6 Diameter Measurements The Eastern Cooperative Oncology Group is one of the national cooperative groups that established carefully
designed response criteria.14 Brain tumor response is measured on computer tomography or MRI by measuring the greatest diameter of the tumor cross section and
MRI measurement of brain tumor response ● L.P. CLARKE
ET AL.
275
physicians then measured diameters on the preselected image slice.
Fig. 2. Choosing the largest diameter is ambiguous, and variability between operators will exist.
its largest perpendicular diameter. The area of the lesion(s) is obtained by multiplying the greatest diameter by its largest perpendicular diameter. This area (or sum of areas if there are multiple lesions) is then used to classify the response into one of four classes: 1) a complete response (CR) is defined as the disappearance of all evidence of active tumor; 2) a partial response (PR) is defined as at least a 50% decrease in the product of the cross sectional diameters; 3) patients with less than a 50% decrease or a 25% increase in the product of the diameters are considered to have stable disease (SD); and 4) progressive disease (PD) is defined as an increase of more than 25% in the product of the cross sectional diameters, or the appearance of new lesions. Three physicians independently measured the greatest diameter and the largest perpendicular diameter on the cross sectional images. Each was presented with all the films with the MR images on a film formatter. The physicians were asked to evaluate each MR data set by choosing the image slice where the lesion showed the largest diameter, and to use that image to do the measurements. The diameters were measured using a compass and the centimeter-scale printed on the film. Considerable variability should be expected with the method. For example, in Fig. 2 a slice is shown where the largest diameter of a single lesion is present. However, physicians may not make the same choice when deciding the location of the tumor boundary. The variability using this method may be partly caused by the choice of image slice. To evaluate the uncertainty due to the interpretation of a single image, we selected from each patient scan one T1-weighted image slice showing the lesion. Four
2.7 GT Criteria Validation of segmentation results for tumor response assessments requires establishing ‘‘ground truth’’ with respect to the tumor volume changes as can be seen on the MR images. Ideally a surgical procedure would provide the true volume, but a single measurement in time would not provide a means for verification of relative changes in tumor volume.2 GT was therefore established using a custom design interface to display full multispectral MRI images and a transparent overlay of the physician determined GT to allow reproducible hand drawing of tumor tissue for each 2D slice. Physician experts (neuroradiologists) generate GT on each slice through the tumor volumes, a very time consuming task. The variation using this method was proven to be small at less than 5%,23 with the source of variation being uncertainty in the image rather than labeling precision. This method has been successfully used in our laboratory for the last 3 years4 – 6,22 for verifying GT specifically for tumor response measurements. However, the method is still based primarily on visual estimates from the T1weighted contrast enhanced image and may not fully represent absolute tumor size or margins. This method should be distinguished from visual extracted 2D methods described above, because this GT method, although manual, requires precision on the voxel level. 2.8 Comparison of the Methods In the experiments we obtained several measurements as is illustrated in Fig. 3. First, physicians measured diameters using a compass on the films of the studies cases. These measurements were used to calculate the product of the diameters, effectively obtaining an ‘‘area’’ enclosing the tumor. Several physicians conducted the measurements, so that an inter-observer variability for both individual diameter measurements and the area measurements could be calculated. As described in section 2.6, the area measurement of a post-treatment MRI study is then compared to the area measurement of the baseline MRI study, and a response class is assigned based on the percentage change of the area. In this study, we also evaluated the inter-observer variability for the response class (Section 3.4). The knowledge guided system and the manually defined GT both yield tumor volumes in cm3. A comparison of the physician determined ‘‘area’’ measurement (in cm2) can only be done in terms of the ranking of the measurements because they have a different unit. To compare response classes, we define decision boundaries on the change in tumor volume. In previous experiments we found that the variability of segmenting MRI data is
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
276
Fig. 3. Measurements obtained for each MRI data set in the study. The visual assessment is done by several physicians, each measuring the largest diameter and the largest perpendicular diameter. The product of these diameters represent an ‘‘area’’ which forms the basis for the response class on follow-up studies. The knowledge guided system automatically segments the whole tumor, yielding a volume. The ground truth method involves manual segmentation of each pixel, also yielding a volume. These volumes are then used for a response class assignment for follow-up studies.
about 5%.4 Therefore, we propose detection of partial response when the tumor shrinks by more than 10%, or progressive disease when the tumor grows by more than 10%. RESULTS 3.1 Area Measurements Three physicians independently measured the greatest diameter and the largest perpendicular diameter on the original MRI films. The product of the measured diameters for each physician is obtained as shown in Fig. 4. The variability is immediately apparent. The coefficient of variation (CoV) is defined as the percent standard # deviation sX of a set of measurements X of the mean X of X:
periment where the physicians picked the image themselves (p 5 0.44). 3.2 Diameter Measurements Because the second set of measurements were done on corresponding images, the measured diameters can be compared. We found that the average CoV for the largest diameter was 10%, where the average CoV for the largest perpendicular diameter was 12%, where the difference is non-significant (p 5 0.30). For both diameters the
sx CoV(X) 5 100% 3 # X For each of the 17 volumes studied, the CoV over the physician measurements was calculated. The average CoV for area measurements was 16%. To gain insight in the source of this variability, a second experiment was done, where four physicians were asked to estimate the diameters on a pre-selected slice from each patient MRI data set. This eliminates the variability due to the choice of slice. The average CoV over the 17 data sets for the area was 19%, where there was no significant difference with the first ex-
Fig. 4. Area measurements for five patients by three physicians. Each patient case has at least the baseline and one follow-up MRI data set evaluated.
MRI measurement of brain tumor response ● L.P. CLARKE
ET AL.
277
Table 1. Spearman correlation coefficients RS for area measurements with ground truth volumes (p 5 significance)
RS p
Fig. 5. Relationship between the manually determined tumor volume and the area measurements done by three physicians.
average variability was 60.5 cm (corresponding to about 1.6 mm on the film). The average correlation coefficient for the diameter measurements was 0.26 6 0.38, indicating that the diameter measurements are independent, i.e., if a physician overestimated the largest diameter (compared to the average), the other diameter was not necessarily also overestimated. In other words, the estimation of the location of the tumor boundary (e.g., see Fig. 2) appears to be the major source of variability, rather then the choice of image or the characteristics of the enhancement, as the latter would be reflected in larger correlation coefficients. 3.3 Comparison with GT Figure 5 depicts the relationship between the GT volume and the physician area measurements, based on the data in the first experiment where each physician selected the image slice to do the measurements on. There is a substantial variability, but more importantly, the relationship with GT is weak. It is unlikely that the relationship can be modeled as a straight line, because the area measurements have a different dimension than the volumes do. The appropriate test statistic is then the Spearman rank-order correlation coefficient RS, a nonparametric test describing the relation: ‘‘if the GT is larger, then the area measurement is larger as well.’’ The results are listed in Table 1. The aggregate for all physician measurements is RS 5 0.39 (p 5 0.005), which indicates that the area measurements, even taken by multiple physicians, do not represent the tumor volume well. For comparison, the Spearman correlation coefficient for GT with the automatically obtained tumor volume using the hybrid knowledge guided segmentation method is RS 5 0.90 (p , 1025). Because GT data and knowledge guided results both
JG
RM
EH
0.49 0.04
0.28 0.28
0.48 0.05
express the tumor volume, the more powerful correlation coefficient for a straight line through the origin can be calculated: RO 5 0.96, showing that the KG system allows accurate response measurements. This is also illustrated in Fig. 6, where the range of physician measurements is shown in gray, and the GT and knowledge guided volumes are drawn in with a separate axis on the right hand side. Where Fig. 5 represents the overall relationship between tumor volume and measured area, Fig. 6 shows that relationship on a patient-by-patient basis. 3.4 Response Measurements In the tumor response measurement protocol, the area measurements are used to establish a response class. Although there is little correlation between area measurements and GT volume, the area measurements may still be adequate for estimation of the response class. Comparison with the volumes found through manual labeling or the knowledge based approach is not straightforward, as the measures represent a different dimension. For purposes of evaluation, we classified a volume measurement as Partial Response when the tumor volume
Fig. 6. Comparisons of response patterns for physician determined areas, knowledge based segmentation and manually determined ground truth volume. The scale for area measurements is on the left y-axis, for volumes the scale is on the right y-axis.
Magnetic Resonance Imaging ● Volume 16, Number 3, 1998
278
Table 2. Response measurements for each follow up (FU) Patient Patient FU2 FU3 FU4 Patient Patient Patient FU2 Patient FU2 FU3 FU4
1, FU1
2, FU1 3, FU1 4, FU1 5, FU1
GT
KG
JG
RM
EH
SD SD PR SD PD PR SD PD PR PR SD PR
SD SD SD SD PD PR SD PD PR PR SD PR
SD SD SD SD SD SD SD PD SD SD SD SD
SD SD SD SD SD SD SD PD PR PR SD PR
SD SD SD SD PD SD SD PD SD SD SD SD
GT 5 ground truth; KG 5 knowledge guided derived volume; JG,RM and EH: response based on physician area measurements; SD 5 stable disease; PR 5 partial response; PD 5 progressive disease.
reduced by more than 10% compared to the baseline image, and we classified the volume measurement as Progressive Disease when the tumor volume increased by more than 10% compared to the baseline. These percentages are smaller than those used for the area measurements, but are reasonable because we found variabilities in tissue volumes on volunteers, when they were imaged serially over similar periods as the brain tumor patients were imaged, to be less than 5%.4 A 10% threshold for the classification is therefore a prudent but warranted decision boundary. As can be seen in Table 2, the three physicians agreed in eight cases on the response, seven of which were ‘‘stable disease’’ according to the criteria for area measurements. However, in only six of those eight times, the area based classification corresponded to the GT volumebased class. In contrast, the knowledge guided segmentation agreed on 11 cases with the GT volume, where for the one case of disagreement the difference in tumor response was only 4%. DISCUSSION AND CONCLUSIONS In this work, we have shown that the protocol for measuring tumor response as commonly used in clinical trials for various cooperative groups does not allow accurate or even correct estimation of the tumor response as can be measured using the gadolinium enhancement in T1-weighted images. The variability in the measurement of the diameters is not the main source of error, because it is only about 10%, which compounds into variability in area measurements of 16%. It is the hypothesis that is the basis for the protocol—that the largest area represents tumor size—that we found to be untrue. At the same time, we have shown that a previously reported auto-
matic segmentation,18 –21 which does not require any operator or physician input, strongly correlates with the tumor volume and, hence, is a viable candidate to replace the area measurements. Acknowledgments–The authors thank Drs. E. Harris, D. Stallworth and G. Fernandez for their help in the film reading and diameter measurements. This work was supported in part by a grant from the National Cancer Institute (R01-CA5942501).
REFERENCES 1. Chaney, E.L.; Pizer, S.M. Defining anatomical structures from medical images. Semin. Radiat. Oncol. 2:215–225; 1992. 2. Clarke, L.P.; Velthuizen, R.P.; Camacho, M.; Heine, J.J.; Vaydianathan, M.; Hall, L.O.; Thatcher, R.W.; Silbiger, M.L. Review of M.A., MRI segmentation: Methods and application. Magn. Reson. Imaging 12:342–368; 1995. 3. Johnston, B.; Atkins, M.S.; Mackiewich, B.; Anderson, M. Segmentation of multiple sclerosis lesions in intensity corrected multispectral MRI. IEEE Trans. Med. Imaging 15: 154 –169; 1996. 4. Vaidyanathan, M.; Heidtman, C.; Clarke, L.P.; Hall, L.O., et al. Normal brain tissue volume measurements using multispectral MRI segmentation methods. Magn. Reson. Imaging 15:87–97; 1997. 5. Velthuizen, R.P.; Clarke, L.P.; Phuphanich, S.; Hall, L.O.; Bensaid, A.M.; Arrington, J.A.; Greenberg, H.M.; Silbiger, M.L. Unsupervised measurement of brain tumor volume on MR images. J. Magn. Reson. Imaging 5:594 – 605; 1995. 6. Vaidyanathan, M.; Clarke, L.P.; Hall, L.O.; Heidtman, C.; Velthuizen, R.P.; Gosche, K.; Phuphanich, S.; Wagner, H.; Greenburg, H.; Silbiger, M.L. Monitoring brain tumor response to therapy using MRI segmentation. Magn. Reson. Imaging 3:323–334; 1997. 7. Sonka, M.; Tadikonda, S.K.; Collins, S.M. Knowledgebased interpretation of MR brain images. IEEE Trans. Med. Imaging 15:443– 452; 1996. 8. Li, X.; Bhide, S.; Kabuka, M.R. Labeling of MR brain images using boolean neural network. IEEE Trans. Med. Imaging 15:628 – 638; 1996. 9. Amit, Y. Graphical shape templates for automatic anatomy detection with applications to MRI brain scans. IEEE Trans. Med. Imaging 16:28 – 40; 1997. 10. Bensaid, A.M.; Hall, L.O. et al. Partially supervised clustering for image segmentation. Pattern Recognition 29: 859 – 871; 1996. 11. Velthuizen, R.P.; Heine, J.J.; Clarke, L.P.; Lin, H. and Fletcher, L.M. Review and Evaluation of MRI nonuniformity corrections for brain tumor response measurements. Medical Physics. In press, 1998. 12. Bensaid, A.M. Improved fuzzy clustering for pattern recognition with applications to image segmentation. Ph.D. Dissertation. Tampa, FL: University of South Florida; 1994. 13. Zijlstra, E.J.; Taphoorn, M.J.B.; Barkhof, F.; Hoogenraad, F.G.C.; Heimans, J.J.; Valk, J. Radiotherapy response of
MRI measurement of brain tumor response ● L.P. CLARKE
14.
15.
16.
17. 18.
19.
20.
21.
22.
cerebral metastasis quantified by serial MR imaging. J. Neuro-Oncology 21:171–176; 1994. Grossman, S.A.; Burch, P.A. Quantitation of tumor response to anti-neoplastic therapy. Semin. Oncol. 15:441– 454; 1988. Watling, C.J.; Lee, D.H.; Macdonald, D.R.; Cairncross, J.G. Corticosteroid-induced magnetic resonance imaging changes in patients with recurrent malignant glioma. J. Clin Oncol. 12:1886 –1889; 1994. Buckner, J.C.; Brown, L.D.; Kugler, J.W.; Cascino, T.L.; Krook, J.E.; Mailliard, J.A.; Kardinal, C.G.; Tschetter, L.K.; O’Fallon, J.R.; Scheithauer, B.W. Phase II evaluation of recombinant interferon alpha and bcnu in recurrent glioma. J. Neurosurg. 82:430 – 435; 1995. Laperrire, N.J.; Berstein, M. Radiotherapy for brain tumors. Cancer J. Clin. 4:96 –108; 1994. Li, C.; Goldgof, D.B.; Hall, L.O. Automatic segmentation and tissue labeling of MR brain images. IEEE Trans. Med. Imaging 12:740 –750; December 1993. Clark, M.C.; Hall, L.O.; Goldgof, D.B.; Clarke, L.P.; Velthuizen, R.P.; Silbiger, M.S. MRI segmentation using fuzzy clustering techniques. IEEE Engineering in Medicine and Biology Magazine 13:730 –742; 1994. Clark, M.C.; Hall, L.O.; Goldgof, D.B. Using fuzzy information in knowledge guided segmentation of brain tumors. In: Fuzzy Logic in AI Workshop at IJCAI, Montreal, CA; August 1995. Clark, M.C.; Hall, L.O.; Goldgof, D.B.; Velthuizen, R.P.; Silbiger, M.S. Automatic tumor segmentation using knowledge-based techniques. IEEE Trans. Med. Imaging In press, 1998. Vaidyanathan, M.; Clarke, L.P.; Velthuizen, R.P.; Phuphanich, S.; Bensaid, A.M.; Hall, L.O.; Bezdek, J.C.; Silbiger, M. Comparison of supervised MRI segmentation
23.
24.
25.
26.
27.
28.
29.
30.
31.
ET AL.
279
methods for tumor volume determination during therapy. Magn. Reson. Imaging 13:719; 1995. Velthuizen, R.P.; Clarke, L.P. An interface for validation of MR image segmentation. In: Proceedings of the 16th Annual International Conference of the IEEE Engineering Medicine and Biology Society; 1994:547–548. Hall, L.O.; Bensaid, A.; Clarke, L.P.; Velthuizen, R.P.; Silbiger, M.L.; Bezdek, J.C. A comparison of neural network and fuzzy clustering techniques in segmenting magnetic resonance images of the brain. IEEE Trans. Neural Networks 3:672– 682; 1992. Bensaid, A.M.; Hall, L.O.; Bezdek, J.C.; Clarke, L.P. Validity-guided (re)clustering for image segmentation. IEEE Trans. Fuzzy Systems 4:112–123; 1996. Kamber, M.; Shingal, R.; Collins, D.L.; Francis, G.S.; Evans, A.C. Model-based 3D segmentation of multiple sclerosis lesions in magnetic resonance brain images. IEEE Trans. Med. Imaging 14:442– 453; 1995. Thompson, P.; Toga, A.W. A surface based technique for warping three-dimensional images of the brain. IEEE Trans. Med. Imaging 15:402– 417; 1996. Sandor, S.; Leahy, R. Surface-based labeling of cortical anatomy using a deformable atlas. IEEE Trans. Med. Imaging 16:41–54; 1997. Galloway, R.L.; Maciunas, R.J.; Failinger, A.L. Factors affecting perceived tumor volumes in magnetic resonance imaging. Ann. Biomed. Engineering 21:367–375; 1993. Bronen, R.; Sze, G. Magnetic resonance imaging contrast agents: Theory and application to the central nervous system. J. Neurosurgy 73:820 – 839; 1990. Bradley, W.G.; Yuh, W.T.C.; Bydder, G.M. Use of MR imaging contrast agents in the brain. J. Magn. Reson. Imaging 3:199 –218; 1993.