Computer-extracted MR imaging features are associated with survival in glioblastoma patients

Computer-extracted MR imaging features are associated with survival in glioblastoma patients Maciej A. Mazurowski, Ph.D. 1, Jing Zhang, Ph.D. 1, Katherine B. Peters, M.D., Ph.D. 2, Hasan Hobbs, M.D. 1 Affiliation: 1. Duke University Medical Center Department of Radiology 2301 Erwin Road, Box 3808 Durham, NC 27710 2. Duke University School of Medicine Department of Neurology Durham, NC 27710 Corresponding Author: Maciej A. Mazurowski, Ph.D Email: maciej.mazurowski@duke.edu Phone: (919) 684-1466 Fax: (919) 684-1491 NOTE: This is the accepted version of the article prior to proof editing. The final version of the article can be found at http://dx.doi.org/10.1007/s11060-014-1580-5 1

ABSTRACT Background: Automatic survival prognosis in glioblastoma (GBM) could result in improved treatment planning for the patient. The purpose of this research is to investigate the association of survival in GBM patients with tumor features in pre-operative magnetic resonance (MR) images assessed using a fully automatic computer algorithm. Methods: MR imaging data for 68 patients from two US institutions were used in this study. The images were obtained from the Cancer Imaging Archive. A fully automatic computer vision algorithm was applied to segment the images and extract 8 imaging features from the MRI studies. The features included tumor side, proportion of enhancing tumor, proportion of necrosis, T1/FLAIR ratio, major axis length, minor axis length, tumor volume, and thickness of enhancing margin. We constructed a multivariate Cox proportional hazards regression model and used a likelihood ratio test to establish whether the imaging features are prognostic of survival. We also evaluated the individual prognostic value of each feature through multivariate analysis using the multivariate Cox model and univariate analysis using univariate Cox models for each feature. Results: We found that the automatically extracted imaging features were associated with survival (p=0.031). Multivariate analysis of individual features showed that two individual features were predictive of survival: proportion of enhancing tumor (p=0.013), and major axis length (p=0.026). Univariate analysis indicated the same two features as significant (p=0.021, and p=0.017 respectively). Conclusion: Computer-extracted MR imaging features can be used for survival prognosis in GBM patients. Key words: Glioblastoma, MRI, survival, computer vision 2

Introduction Glioblastoma (GBM) is the most common [1] and the most aggressive primary brain tumor [2, 3]. The 1-year survival rate is 50.6% [3], and the 5-year survival rate is 9.8% [2]. While the prognosis is usually poor in this population, there is a subset of patients that experience a better prognosis and can have long-term survival (as defined by > 5 years from time of diagnosis). In light of these differences, biomarkers for survival prognosis could be of high utility. One of the benefits of such biomarkers is the improved selection of patients for clinical trials. Different variables have been shown to be associated with survival in GBM. Clinical features predictive of time to death include Karnofsky Performance Status (KPS) [4, 5] and age [4]. Extent of resection is also associated with survival such that gross total resection is associated with a better prognosis [4]. Recently, prognostic and predictive biomarkers for GBM based on genetic information have been developed [6]. Examples include a gene expression-based tumor classification by Verhaak et al [7] which was shown to be associated with patient survival and response to aggressive therapy, as well as a multi-gene biomarker for survival prognosis presented by Colman et al [8]. Mutations in isocitrate dehydrogenase 1 (IDH1) [9] and methylguanine methyl transferase (MGMT) promoter methylation [10-12] are now commonly tested and used prognostically but it remains to be seen if they can be predictive for survival time in patients with GBM. Prognostic biomarkers based medical imaging have been explored to a lesser extent. However, the research in this direction is gaining momentum. Lacroix et al [4] has shown that 4 out of 7 of the examined imaging features were associated with survival. These features were: tumor functional grade, necrosis grade, edema grade, and enhancement grade. Studies by Pope et al [13], Park et al [5], and Zinn et al [14] provided more evidence for the usefulness of imaging features when predicting survival time. Recently, two studies by Gutman et al [15], and Mazurowski et al [16] used a standardized lexicon of imaging features extracted from MRI scans called VASARI and showed that such features are predictive of survival. Survival prognosis through the utilization of standard of care preoperative imaging would be of value as it captures phenotypical properties of the tumor that might not be currently fully identified by other biomarkers. Furthermore, it does not require conducting any additional tests as the imaging information is already available through standard of care imaging. The disadvantage is a significant additional workload for radiologists and therefore additional cost of 3

the test incurred by the time consuming feature assessment. Furthermore, the interobserver variability [15] in the feature assessment could negatively affect the usefulness of such features. In this study, we address this shortcoming of medical imaging biomarkers for GBM and propose a biomarker for survival prognosis that is based on MR imaging features extracted automatically using computer algorithms. Since the imaging features are extracted in a fully automatic manner and the survival prognosis is done using a computer model, such survival prognosis could be provided automatically to the clinician interpreting/ordering the imaging exam. This study is a proof of concept for such an approach. Specifically, in this pilot study, we automatically extracted 8 features and demonstrated that they were associated with survival. Additional benefits of a computerized biomarker are a lack of interobserver variability (i.e. the algorithm always provides the same response to a given image) and precision of the measurements. This is the first study that demonstrates that fully automatically extracted MR imaging features are predictive of survival. However, there are some results in the literature that are related to our finding. The closest one is in the paper by Zacharaki et al [17], where the authors examined a decision tree for predicting survival that used manually extracted features, and semiautomatically extracted features (tumor segmentation was aided by experts). In their univariate analysis, the authors showed a relationship between some simple intensity-based features (such as mean intensity) of two tumor compartments distinguished by the authors observed in DTI for high grade glioma patients (a portion were GBM patients). Another very recent study [18] applied automatically extracted MR imaging features for survival prediction in 16 patients but did not provide statistical analysis that allowed a determination of the association between those features and survival. Multiple other studies, for example [19],[20], and [21] present computer algorithms for the segmentation of brain tumors (some of them into different compartments) but those studies do not offer an evaluation of such methods in the context of survival prognosis. 4

Materials and Methods Patient Population Our patient population consisted of 68 GBM patients. Out of the 68 patients, 22 were from MD Anderson Cancer Center and 46 were from Henry Ford Hospital. For each patient, preoperative MR imaging data as well as survival data were available. The patient survival data was obtained from the Cancer Genome Atlas (TCGA, http://cancergenome.nih.gov/). The two databases are correlated through unique identifiers. The list of identifiers for patients used in our study can be accessed at the following link: http://deckard.duhs.duke.edu/~mazurowski/data/tcga-gbm-68- patients-list.txt. Imaging Data The patient imaging data was downloaded from the Cancer Imaging Archive (TCIA, http://cancerimagingarchive.net/). For each patient, three MR imaging sequences were available. The sequences were: (1) T1-weighted pre contrast sequence, (2) T1-weighted post contrast sequence, and (3) FLAIR. More information about the imaging data can be found on the TCIA website (http://cancerimagingarchive.net/). The images were preoperative. We annotated different tumor components in each slice where a tumor occurs. All annotations were approved by a neuroradiology fellow at our institution. The annotations were used to train our segmentation algorithm. Algorithm for extracting imaging features The imaging features used in our study were extracted using a computer algorithm developed in our laboratory and presented in another, accompanying publication [22] which focuses on the details of the algorithm, comparison of the automatic and manual segmentations, and comparison between the values of the extracted features and their reference standard. The first step of the algorithm was automatic segmentation of the tumor into four mutually exclusive compartments: (C1) necrosis, (C2) enhancing tumor, (C3) T1 abnormality excluding C1 and C2, (C4) FLAIR abnormality excluding T1 abnormality. Fig. 1 shows the four tumor compartments. The segmentation was done by applying a random forest classifier to each voxel in order to classify it into one of the four tumor compartment (or as normal). The classifier used 219 low-level descriptors extracted for each voxel based on the three available imaging sequences (the 5

sequences were registered prior to segmentation). More details on this algorithm can be found in [22], where we evaluate the performance of the segmentation algorithm in detail including measurements of agreement between manual and our automatic segmentations using Dice coefficient. After the tumors were identified in the images and segmented into different compartments, high level imaging features were extracted as described below. Fig. 1 Tumor segmentation into different compartments: C1 : enhancing region; C2 : necrotic region; C3 : T1 abnormality (hypo-intensity region on T1) excluding enhancing and necrotic regions; C4 : FLAIR abnormality (hyper-intensity region on FLAIR) excluding T1 abnormality. Figure (A) shows the manually annotated compartments. Figure (B) shows the compartments as classified by our segmentation algorithm. Figure (C) shows the original post contrast T1-weighted slice. Imaging features Based on the tumor segmentation provided by our algorithm described above, we extracted 8 imaging features for each case. These were features that we could reliably extract based on our automatic segmentation algorithm and were potentially predictive of survival. These features were: tumor side, proportion of enhancing tumor, proportion of necrosis, T1/FLAIR ratio, major axis length, minor axis length, tumor volume, and thickness of enhancing margin. The proportion of necrosis and proportion of enhancing tumor were calculated as the volumes of the respective compartments divided by the volume of the sum of all compartments. The T1/FLAIR ratio was calculated as the proportion of the total volume of tumor compartments C1, C2, and C3, and the total volume of all tumor compartments. The volume, major axis length, and minor axis length 6

were calculated for the entire tumor understood as the sum of all compartments. The thickness of the enhancing margin was assigned a value of 0 if the tumor was solidly enhancing. Experimental design and statistical analysis First, we applied a well-known leave-one-out cross validation approach to automatically extract imaging features. Specifically, we used all but one case to train our feature extraction algorithm which involved a random forest classifier. Then, we applied the trained algorithm to extract the 8 imaging features for the one case that was left out. This was repeated 68 times until the features were extracted for each case. This approach was applied to avoid a possible bias caused by using the same images for training and testing of the algorithm. The features extracted in this way were then used for the survival analysis as described below. To test the main hypothesis of the study, we developed a Cox proportional hazards regression model including all imaging features as independent variables. A likelihood ratio test was used to establish whether there was a significant association between these features and overall survival. Specifically, the coxph function in R was used for this purpose. We also explored the association of individual features with survival. For multivariate analysis, we used the same Cox proportional hazards regression model to establish the significance of each individual feature. For the univariate analysis, we constructed a univariate Cox proportional hazards regression model for each variable individually and conducted a likelihood ratio test to establish whether there was a significant association between that variable and survival (R s coxph function was used). As an additional exploratory analysis, we used two imaging features indicated as significant in the analysis to divide the population of patients into four groups and we evaluated survival in each of these groups. Toward this goal, we binarized each variable such that the values higher than the median value of the variable for the population were changed to high and values lower or equal than median were changed to low. Then, four possible groups were established for the combinations of the values for the two variables. We plotted the survival curves for each of the groups and ran the log-rank test to compare the curves. Specifically, we used the survdiff function (with rho=0) in R for this purpose. 7

Results The distribution of values for each of the imaging features is presented on the histograms in Fig. 2. As expected, one can see generally higher values for the proportion of enhancement than for the proportion of necrosis. Even higher values can be observed for the T1/FLAIR ratio. The major axis length of the tumor was generally under 10 cm and the volume was under 200 cm 3. Table 1: Patient characteristics as well as follow up times. For continuous or semi-continuous variables, we provide mean and range. Characteristic Patients (n=68) Age at diagnosis (years) 59 (18-84) Sex Male 44 Female 24 Race/Ethnicity White 59 Black or African American 3 Asian 5 Not available 1 Follow up (days) 484 (16-1757) Deaths recorded 55 Regarding the main hypothesis of our study, the multivariate analysis showed that the automatically extracted image features were predictive of survival (p=0.031). Exploratory multivariate analysis for individual features indicated that two features were associated with survival: proportion of enhancing tumor (p = 0.013) and major axis length (p = 0.026). The univariate analysis confirmed these findings by indicating that the same two variables were associated with survival: proportion of enhancing tumor (p=0.021) and major axis length (p = 8

0.016). None of the other variables reached significance at the 0.05 level in the univariate or multivariate analyses. Fig. 2 Histograms of the extracted MR imaging features for the 68 study patients. Horizontal axes show feature values. For the variable Side, 1=Left, 2=Center/Bilateral, 3=Right. As additional preliminary analysis, we used the proportion of enhancing tumor and major axis length to divide our cohort of patients into four groups to establish whether any of these groups were associated with a worse or improved survival. The survival curves for each of the groups are presented in Fig. 3. One can see that the group with low major axis length and small proportion of enhancing tumor showed notably improved survival. The log rank test indicated that the survival curves in Fig. 3, were significantly different (p= 0.00759). 9

Fig. 3 Kaplan-Mayer curves for four patient groups determined based on major axis length and proportion of enhancing tumor Discussion In this paper we propose that computer-extracted preoperative MR imaging features can be used to predict survival time in GBM patients. We conducted an experiment with 68 subjects for whom we automatically extracted 8 imaging features and demonstrated that these features are associated with survival. The proportion of enhancing tumor and major axis length stood out as significant predictors of survival in multivariate and univariate analysis. The association of major 10

axis length with survival found in this study, while will require additional validation, adds to only limited prior data on importance of this feature [15, 16]. We believe that the proposed approach is a significant step toward a more systematic use of MR imaging in GBM survival prognosis. The proposed approach is very practical in that it does not require additional work from the radiologist interpreting the image. The advantages of our approach are the following. First, it relieves a radiologist from the burden of assessing multiple additional features in the imaging study. This would decrease the cost of a prognostic biomarker based on our approach. Second, we offer a quantitative approach to feature assessment. Specifically, the volumetric measurements are performed precisely using computer algorithms and do not rely on an approximate assessment by a radiologist. Related to this, our approach is free of inter-observer variability. Finally, while in this study we have evaluated only simple features, in general the computer-based approach allows for the assessment of a large number of features, some of which might not be easily perceptible to human eye. Examples include texture features. However, additional experiments are necessary to evaluate the prognostic value of such features. We envision that the feature extraction and survival prediction software would be embedded into the imaging equipment and the survival prognosis would be provided automatically to the clinician along with the actual imaging study. Our study provides a proof of concept that the utilization of GBM MR images in such manner is possible. While our study provides a proof of concept for the automatic approach survival prognosis based on MR imaging, significant further development is needed to improve such prediction. Specifically, our study is limited to only 8 relatively simple features. A multitude of additional features could be extracted (e.g. texture features), some of which might be related to survival. The evaluation of such additional features will require study with a larger number of patients. Regarding the two features that were shown to be associated with survival in our study, the proportion of enhancing tumor and major axis length of the tumor, our finding is consistent with previously reported findings where manually extracted equivalents of these features were found to be strong predictors of survival [15]. Our additional exploratory analysis of the model with just these two variables showed that a group associated with low major axis length and low proportion of enhancing tumor is associated with a notable improvement in survival. While this finding should be confirmed with more analysis including a larger database of patients, it shows potential for a simple automated imaging biomarker for survival prognosis. 11

Our experiments showed some robustness of our approach. Specifically, in the evaluation process, we included data from two different institutions with varying presentations of different compartments in the MR images. Such variability makes the automatic segmentation of different tumor compartments more difficult. However, regardless of this variability, the features extracted by our algorithm remained significant predictors of survival. Our study has some limitations. One of them is a limited sample size (n=68). A larger sample size could unveil additional variable significant in predicting survival. Please note, however, that the sample size used in our study is not unusual for this kind of investigation. Please also note that our study required annotation of four different tumor compartments of each slice of each case which required enormous amount of radiologist work. We are not aware of a study using similar data for 68 patients. Furthermore, our proof-of-concept study does not include clinical and treatment variables such as age, Karnofsky performance status, or type of surgery. Future studies with larger sample size will include such variables. To conclude, in this study, we presented a proof of concept for using computer-extracted imaging features to predict survival in GBM. While further studies are necessary, our approach was shown to be promising. Conflict of Interest: Dr. Maciej A. Mazurowski receives grant funding from the Department of Defense Breast Cancer Research Program. He also receives consulting fees from American College of Radiology Image Metrix (contractor to GE) for his services as a scientific consultant. Jing Zhang, Katherine B. Peters, and Hasan Hobbs have no conflicts of interest to declare. REFERENCES 1. Dolecek TA, Propp JM, Stroup NE, Kruchko C. CBTRUS statistical report: primary brain and central nervous system tumors diagnosed in the United States in 2005 2009. Neuro- Oncology. 2012;14(suppl 5):v1-v49. 2. Stupp R, Hegi ME, Mason WP, et al. Effects of radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone on survival in glioblastoma in a randomised phase III study: 5-year analysis of the EORTC-NCIC trial. The lancet oncology. 2009;10(5):459-66. 12

3. Stupp R, Mason WP, Van Den Bent MJ, et al. Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. New England Journal of Medicine. 2005;352(10):987-96. 4. Lacroix M, Abi-Said D, Fourney DR, et al. A multivariate analysis of 416 patients with glioblastoma multiforme: prognosis, extent of resection, and survival. Journal of neurosurgery. 2001;95(2):190-8. 5. Park JK, Hodges T, Arko L, et al. Scale to predict survival after surgery for recurrent glioblastoma multiforme. Journal of Clinical Oncology. 2010;28(24):3838-43. 6. McDonald KL, Aw G, Kleihues P. Role of biomarkers in the clinical management of glioblastomas: What are the barriers and how can we overcome them? Frontiers in neurology. 2012;3. 7. Verhaak RG, Hoadley KA, Purdom E, et al. An integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR and NF1. Cancer cell. 2010;17(1):98. 8. Colman H, Zhang L, Sulman EP, et al. A multigene predictor of outcome in glioblastoma. Neuro-oncology. 2010;12(1):49-57. 9. Yan H, Parsons DW, Jin G, et al. IDH1 and IDH2 mutations in gliomas. New England Journal of Medicine. 2009;360(8):765-73. 10. Gorlia T, van den Bent MJ, Hegi ME, et al. Nomograms for predicting survival of patients with newly diagnosed glioblastoma: prognostic factor analysis of EORTC and NCIC trial 26981-22981/CE. 3. The lancet oncology. 2008;9(1):29-38. 11. Rivera AL, Pelloski CE, Gilbert MR, et al. MGMT promoter methylation is predictive of response to radiotherapy and prognostic in the absence of adjuvant alkylating chemotherapy for glioblastoma. Neuro-oncology. 2010;12(2):116-21. 12. Minniti G, Salvati M, Arcella A, et al. Correlation between O6-methylguanine-DNA methyltransferase and survival in elderly patients with glioblastoma treated with radiotherapy plus concomitant and adjuvant temozolomide. Journal of neuro-oncology. 2011;102(2):311-6. 13. Pope WB, Sayre J, Perlina A, Villablanca JP, Mischel PS, Cloughesy TF. MR imaging correlates of survival in patients with high-grade gliomas. American journal of neuroradiology. 2005;26(10):2466-74. 14. Zinn PO, Sathyan P, Mahajan B, et al. A novel volume-age-kps (VAK) glioblastoma classification identifies a prognostic cognate microrna-gene signature. PLoS One. 2012;7(8):e41522. 15. Gutman DA, Cooper LA, Hwang SN, et al. MR Imaging Predictors of Molecular Profile and Survival: Multi-institutional Study of the TCGA Glioblastoma Data Set. Radiology. 2013. 13

16. Mazurowski MA, Desjardins A, Malof JM. Imaging descriptors improve the predictive power of survival models for glioblastoma patients. Neuro-oncology. 2013;15(10):1389 94. 17. Zacharaki E, Morita N, Bhatt P, O'Rourke D, Melhem E, Davatzikos C. Survival Analysis of Patients with High-Grade Gliomas Based on Data Mining of Imaging Variables. American Journal of Neuroradiology. 2012;33(6):1065-71. 18. Zhou M, Hall LO, Goldgof DB, Gillies RJ, Gatenby RA. Survival time prediction of patients with glioblastoma multiforme tumors using spatial distance measurement. SPIE Medical Imaging: International Society for Optics and Photonics, 2013; p. 86702O-O-7. 19. Popuri K, Cobzas D, Murtha A, Jägersand M. 3D variational brain tumor segmentation using Dirichlet priors on a clustered feature set. International journal of computer assisted radiology and surgery. 2012;7(4):493-506. 20. Zhu Y, Young GS, Xue Z, et al. Semi-automatic segmentation software for quantitative clinical brain glioblastoma evaluation. Academic Radiology. 2012;19(8):977-85. 21. Zikic D, Glocker B, Konukoglu E, et al. Decision forests for tissue-specific segmentation of high-grade gliomas in multi-channel MR. Medical Image Computing and Computer-Assisted Intervention MICCAI 2012: Springer, 2012; p. 369-76. 22. Zhang J, Barboriak DP, Hobbs H, Mazurowski MA. A fully automatic extraction of magnetic resonance image features in Glioblastoma patients. Medical Physics. 2013; 41(4):042301. 14