FULL PAPERS

Magnetic Resonance in Medicine 000:000000 (2011)

Nonlinear Registration of Diffusion-Weighted Images Improves Clinical Sensitivity of Functional Diffusion Maps in Recurrent Glioblastoma Treated With Bevacizumab
Benjamin M. Ellingson,1* Timothy F. Cloughesy,2 Albert Lai,2 Phioanh L. Nghiemphu,2 and Whitney B. Pope1
Diffusion-weighted imaging estimates of apparent diffusion coefficient (ADC) have shown sensitivity to brain tumor cellularity as well as response to therapy. Functional diffusion maps (fDMs) exploit these principles by examining voxelwise changes in ADC within the same patient over time. Currently, the fDM technique involves linear image registration of ADC maps from subsequent follow-up times to pretreatment ADC maps; however, misregistration of ADC maps due to geometric distortions as well as mass effect from growing tumor can confound fDM measurements. In this study, we compare the use of a nonlinear registration scheme to the current linear fDM technique in 70 patients with recurrent glioblastoma multiforme treated with bevacizumab. Results suggest that nonlinear registration of pretreatment ADC maps to posttreatment ADC maps improves the clinical predictability, sensitivity, and specificity of fDMs for both progression-free and C overall survival. Magn Reson Med 000:000000, 2011. V 2011 Wiley-Liss, Inc. Key words: functional diffusion map; fDM; glioblastoma; bevacizumab; diffusion MRI

Glioblastoma multiforme (GBM), a particularly infiltrative and aggressive type of glioma, makes up nearly 17% of all primary brain tumors and approximately 54% of all primary brain and central nervous system gliomas (1). GBMs are trademarked by a very poor patient prognosis, with a mean survival of only around 14.6 months following standard therapy including radiation and chemotherapy (2). Because of this poor prognosis, there is a great need for early imaging biomarkers to evaluate effectiveness of new therapies and predict patient survival. Diffusion-weighted magnetic resonance imaging (DWI) techniques have shown value in evaluating regions suspected of tumor invasion and proliferation (3), as well as the efficacy of specific treatment paradigms. Through calculation of an apparent diffusion coefficient (ADC) from DWIs, multiple studies have reported an inverse relationship

1 Department of Radiological Sciences, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, California, USA. 2 Department of Neurology, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, California, USA. Grant sponsors: Brain Tumor Funders Collaborative, Art of the Brain, Ziering Family Foundation in memory of Sigi Ziering, Singleton Family Foundation, Clarence Klein Fund for Neuro-Oncology. *Correspondence to: Benjamin M. Ellingson, Ph.D., Department of Radiological Sciences, David Geffen School of Medicine, University of California Los Angeles, 924 Westwood Boulevard, Suite 615, Los Angeles, CA 90024. E-mail: bellingson@mednet.ucla.edu Received 25 January 2011; revised 21 April 2011; accepted 24 April 2011. DOI 10.1002/mrm.23003 Published online in Wiley Online Library (wileyonlinelibrary.com).
C V 2011 Wiley-Liss, Inc.

between ADC and tumor cell density (46), suggesting that DWIs can potentially be used as imaging biomarkers to evaluate growing tumor and treatment effectiveness. The functional diffusion map (fDM) was first proposed in 2005 as a method to detect changes in ADC known to accompany successful cytotoxic therapies (7,8). In contrast to other techniques that average the ADC measurements for an entire region, fDMs do not assume homogeneity within tumors. Instead, multiple ADC maps collected on different days are coregistered with a baseline image at an initial time point, and then regional (voxelwise) changes in ADC are isolated and labeled according to the magnitude of their change. This technique has shown to be one of the most sensitive early biomarkers for tumor response to chemotherapeutics and radiotherapy and has shown to be highly specific to progression of high-grade gliomas (9,10). Additionally, investigators have shown that fDMs can be used as a tool for long-term clinical assessment of recurrent brain tumors (11), are sensitive to infiltrating tumors such as gliomatosis cerebri (12), and allow visualization and quantification of regional changes in cellularity that correlate with tumor status and response to various treatments (11,13). Despite promising results from early fDM studies, proper image registration between DWI datasets remains a major limitation to this technique. Slight misregistration between DWI datasets due to image distortion as well as mass effect from edema or growing tumor can potentially confound the interpretation and quantification of fDM-classified tumor regions. To date, implementation of fDMs for brain tumor evaluation has consisted of only linear image registration techniques to match baseline DWIs to subsequent DWI datasets. We hypothesize that an additional step of nonlinear (elastic) registration after a linear registration step may improve the clinical sensitivity of fDMs when applied to the brain, reducing the effects of misregistration and misclassification. In this study, we explore the effects of nonlinear registration of DWIs on fDMs in patients with recurrent glioblastoma before and after treatment with the antiangiogenic agent bevacizumab. MATERIALS AND METHODS Patients All patients participating in this study signed institutional review board-approved informed consent to have
1

Ellingson et al.

their data collected and stored in our institutions neurooncology database. Data acquisition was performed in compliance with all applicable Health Insurance Portability and Accountability Act regulations. The study spanned November 15, 2005 to August 31, 2010. Patients were retrospectively selected from our institutions neurooncology database. A total of n 252 patients who met the following criteria were initially selected: (1) pathology confirmed GBM with recurrence based on MRI and clinical data, (2) regularly treated every 2 weeks per cycle with bevacizumab (Avastin, Genentech, South San Francisco, CA; 5 or 10 mg/kg body weight), and (3) baseline (pre-bevacizumab treatment) and minimum of one followup MRI scans. Of these patients, n 70 patients had good quality DWIs before and after initiation of bevacizumab treatment and had adequate image registration (linear and nonlinear) without gross misalignment. Baseline scans were obtained approximately 1.5 weeks before treatment, and follow-up scans were obtained at approximately 6 weeks after the initiation of bevacizumab. At the time of last assessment (August, 2010), 57 of the 70 patients were deceased. All patients were treated with radiation therapy (typically 6000 cGy) and maximal tumor resection at time of initial tumor presentation. Patients did not receive additional radiation or chemotherapy between pretreatment and post-treatment MRI scans. MRI Data were collected on a 1.5-T MR system (General Electric Medical Systems, Waukesha, WI) using pulse sequences supplied by the scanner manufacturer. Standard anatomical MRI sequences included axial T1-weighted (echo time [TE]/repetition time [TR] 15 ms/400 ms, slice thickness 5 mm with 1 mm interslice distance, number of excitations (NEX) 2, matrix size 256 256, and field-of-view [FOV] 24 cm), T2-weighted fast spin-echo (TE/TR 126130 ms/4000 ms, slice thickness 5mm with 1 mm interslice distance, NEX 2, matrix size 256 256, and FOV 24 cm), and fluid-attenuated inversion recovery (FLAIR) images (inversion time 2200 ms, TE/TR 120 ms/4000 ms, slice thickness 5 mm with 1 mm interslice distance, NEX 2, matrix size 256 256, and FOV 24 cm). DWIs were collected with TE/TR 102.2 ms/8000 ms, NEX 1, slice thickness 5 mm with 1 mm interslice distance, matrix size 128 128 (reconstructed images were zero-padded and interpolated to 256 256), and a FOV 24 cm using a twicerefocused spin echo echo planar preparation (14). ADC images were calculated from acquired DWIs with b 1000 s/mm2 and b 0 s/mm2 images. Additionally, gadopentetate dimeglumine-enhanced (Magnevist; Berlex, Wayne, NJ; 0.1 mmol/kg) axial and coronal T1-weighted images (coronal: TE/TR 15 ms/400 ms, slice thickness 3 mm with 1mm interslice distance, NEX 2, a matrix size of 256 256, and FOV 24 cm) were acquired immediately after contrast injection. Linear Registration All images for each patient were registered to baseline anatomical T1-weighted images using a mutual informa-

tion algorithm and a 12-degree of freedom transformation using FSL (FMRIB, Oxford, UK; http://www.fmrib.ox. ac.uk/fsl/). Fine registration (12 degrees and 12 voxels) was then performed using a Fourier transform-based, 6degree of freedom, rigid body registration algorithm (15) followed by visual inspection to ensure adequate alignment. Similar registration techniques were used in previous fDM studies (5,7,12,16). Nonlinear Registration After linear registration was complete, nonlinear registration was implemented using the FMRIB non-linear registration tool (FNIRT), a free-form deformation algorithm using a B-spline basis function (1719). FNIRT uses a Levenberg-Marquardt modification of the Gauss-Newton method for optimizing the minimum of a sum-of-squares cost function to perform the nonlinear registration. Default values for registration parameters supplied by FMRIB were used in this study (see http://www.fmrib. ox.ac.uk/fsl/fnirt/index.html for more details). Two different nonlinear registrations were performed (Fig. 1b and c). First, we performed nonlinear registration on pretreatment ADC maps to register to post-treatment ADC maps (pre-to-post). This registration scheme was thought to reduce the amount of mass effect and edema present in the pretreatment ADC maps by warping to the post-treatment ADC maps. Second, we performed nonlinear registration on post-treatment ADC maps to align with pretreatment ADC maps (post-to-pre). We expected this registration scheme to produce similar fDM measurements to the linear case, as much of the effects are likely to be due to reduction in edema. As FNIRT uses a sum-of-squares cost function, which may lead to problems due to dependence on image intensities, we also tested a nonlinear registration algorithm using the image registration toolkit (IRTK), which uses cubic B-splines basis function and a normalized mutual information cost function (20,21). Default values for registration parameters supplied by IRTK were used (see http://www.doc.ic.ac.uk/$dr/software/usage.html for more details). Additionally, we explored the use of nonlinear registration between T1-weighted anatomical datasets to determine whether results differed compared to registering ADC maps directly. Specifically, nonlinear registration was used to align pretreatment and posttreatment T1-weighted anatomical images, and then this transformation was subsequently applied to link pretreatment and post-treatment ADC maps. Regions of Interest In this study, we chose to examine both regions of FLAIR signal abnormality on pretreatment FLAIR images as well as regions of contrast enhancement on pretreatment, postcontrast T1-weighted images to provide a comprehensive comparison. FLAIR regions of interest (ROIs) (5,1113,22) and contrast-enhancing ROIs (79,16) have both previously been used in interpreting fDM results. For isolating ROIs, we used a semiautomated process consisting of: (1) manually defining the relative region of tumor occurrence, then (2) thresholding either the FLAIR or postcontrast

Nonlinear fDMs in Recurrent GBM

FIG. 1. Linear and nonlinear fDM calculations. a: Linear (traditional) fDMs consisting of using a linear registration algorithm to align post-treatment ADC maps to pretreatment ADC maps. b: Pre-to-post nonlinear fDMs consisting of nonlinear registration of pretreatment ADC maps to post-treatment ADC maps. c: Post-to-pre nonlinear fDMs consisting of nonlinear registration of post-treatment ADC maps to pretreatment ADC maps. For fDMs, blue voxels represent a significant decrease in ADC (beyond 0.4 mm2/ms), red voxels represent a significant increase in ADC (beyond 0.4 mm2/ms), and green voxels are those with no significant change in ADC.

images using an empirical threshold combined with a region-growing algorithm, and then (3) manually editing the resulting masks to exclude obvious errors. Functional Diffusion Map Calculation fDM calculation was performed as recommended in a previous publication (5). After proper registration was visually verified, voxelwise subtraction was performed between acquired, post-treatment ADC maps and baseline, pretreatment ADC maps. Individual voxels were stratified into three categories based on the change in ADC relative to the baseline ADC map. Voxels were defined as significantly increasing or significantly decreasing if DADC values were beyond 0.40 mm2/ms, which has been suggested to be the best threshold for a balance between sensitivity and specificity to progressing disease and corresponds to an estimated change in

cell density of more than 63960 nuclei/mm2 in untreated gliomas (5). The percent fractional volume of increasing (%VADC > 0.4 mm2/ms; red) and decreasing (%VADC < 0.4 mm2/ms; blue) voxels within FLAIR and contrast-enhancing regions were used as biomarkers for comparing linear to nonlinear fDMs and for predicting patient outcomes. Hypothesis Testing We hypothesized that fDMs generated from ADC maps using nonlinear registration (nonlinear fDMs) would have significantly lower volume fraction of changing voxels (increasing or decreasing) compared to fDMs generated using only linear registration (linear fDMs). To test this hypothesis, we performed multiple repeated measures, two-way analysis of variance for each of the fDM measures (%VADC > 0.4 mm2/ms and

Ellingson et al.

%VADC < 0.4 mm2/ms), ROIs (FLAIR and contrast enhancing), and registration algorithms (FNIRT-ADC, FNIRT-Anatomy, IRTK-ADC, and IRTK-Anatomy) across the different registration schemes (linear, pre-to-post nonlinear, and post-to-pre nonlinear). Tukeys test for multiple comparisons was used for post hoc comparisons. Additionally, we tested whether the volume fraction of tissue showing a significant decrease in ADC using either the linear or nonlinear techniques was a significant predictor of time-to-progression (TTP) or overall survival (OS) with respect to the post-treatment MRI scan date. The volume fraction of decreasing ADC (blue regions) was used as a predictor of patient outcomes because it is thought to reflect both reduction in edema as well as a change in tumor cellularity after bevacizumab treatment (13,22). Disease progression was defined by a modified Macdonald criteria (23,24), which included either a 25% increase in enhancing tumor or an increase in nonenhancing tumor evident by increased mass effect and/or architectural distortions such as the blurring of the graywhite matter interface. We hypothesized that nonlinear fDMs would be a better predictor of these outcomes compared to linear fDMs. Log-rank statistical analyses on Kaplan Meier data were performed using GraphPad PrismV v4.0 statistical software to test this hypothesis. Additionally, the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated for 3-month progression-free survival (PFS), 6-month PFS, 6month OS, and 12-month OS using each of the fDM metrics, ROIs, and registration schemes.
R

RESULTS In general, linear fDMs showed a substantial volume of tissue with decreased ADC after treatment with bevacizumab, presumably due to reduction in edema or growing tumor, both of which are thought to reduce ADC within FLAIR abnormal regions (Fig. 1a). When examining pre-to-post nonlinear fDMs, the quantity of tissue with high ADC was typically lower as the pretreatment ADC maps were warped to the post-treatment, edemareduced, ADC maps (Fig. 1b). This resulted in a general decrease in the number of fDM-classified voxels. Qualitatively, post-to-pre nonlinear fDMs appeared similar to linear fDMs, showing a large volume of tissue with a significant reduction in ADC (Fig. 1c) likely due to either reduction in edema or increased tumor cellularity. Quantitative analysis confirmed these observations, suggesting a significant difference in all fDM metrics and all ROIs due to the different registration schemes (Fig. 2; repeated measures, two-way analysis of variance, P < 0.001 for all fDM metrics and ROIs); however, no differences were observed between nonlinear registration algorithms or due to registration of ADC versus anatomical images (repeated measures, two-way analysis of variance, P > 0.8 for all combinations and interactions). Within contrast-enhancing regions, the fractional volume of tissue having a significantly lower ADC (blue regions) was reduced in pre-to-post nonlinear fDMs compared to both linear fDMs in Fig. 2a (Tukey, P < 0.05 for all registration algorithms) and post-to-pre nonlinear fDMs (Tukey, P < 0.05 for all registration algorithms). The fractional

volume of tissue having a significantly higher ADC (red regions) within pretreatment contrast-enhancing regions was significantly different between linear fDMs and post-to-pre nonlinear fDMs (Fig. 2b; Tukey, P < 0.01 for all registration algorithms). Similar to contrast-enhancing ROIs, the fractional volume of tissue having significantly lower ADC within FLAIR ROI was significantly lower in pre-to-post nonlinear fDMs compared to linear fDMs (Fig. 2c; Tukey, P < 0.01 for all registration algorithms). The volume fraction of tissue having significantly increasing ADC within FLAIR ROIs was significantly different between linear fDMs and nonlinear fDMs (Fig. 2d; Tukey, P < 0.001 for linear fDMs vs. both pre-to-post and post-to-pre nonlinear fDMs). Next, we tested the hypothesis that a fractional volume of tissue having a decrease in ADC larger than the median change in all patients results in a shorter TTP and shorter OS, and that nonlinear fDMs provide better stratification of survival compared to linear fDMs. As no statistical difference was observed between volume fractions calculated with FNIRT versus IRTK, or between ADC registration and registration performed between T1weighted anatomical datasets (Fig. 2), we chose to test survival differences using FNIRT applied to ADC maps. Within contrast-enhancing ROIs, the median percentage of tumor volume exhibiting a significant decrease in ADC for linear and pre-to-post nonlinear fDMs was not able to stratify TTP (Fig. 3a; log-rank, linear fDMs, P 0.1060, pre-to-post nonlinear fDMs, P 0.0666) or OS (Fig. 3b; log-rank, linear fDMs, P 0.1229, pre-to-post nonlinear fDMs, P 0.0539). Interestingly, the fractional volume of tissue with decreasing ADC within contrastenhancing ROIs in post-to-pre nonlinear fDMs were able to stratify long- and short-term TTP (log-rank, P 0.0167) and OS (log-rank, P 0.0260), despite not having statistically different fractional volumes to linear fDMs after pairing (Fig. 2a; linear fDMs vs. nonlinear fDMs, P > 0.05). Within FLAIR ROIs, however, both linear and nonlinear fDMs were able to significantly stratify short- and long-term TTP and OS (Fig. 3c and d). Specifically, a volume fraction of tissue having a significant decrease in ADC within FLAIR ROIs larger than the median using linear fDMs resulted in a statistically shorter TTP (Fig. 3c; Log-rank, P 0.0016) and OS (Fig. 3d; Log-rank, P 0.0385). Post-to-pre nonlinear fDMs showed similar trends to linear fDMs with respect to TTP (Log-rank, P 0.0171) and OS (Log-rank, P 0.0137); however, pre-to-post nonlinear fDMs showed a substantial reduction in the level of significance (P value) compared to both linear and post-to-pre nonlinear fDMs for TTP (Log-rank, P 0.0009) and OS (Logrank, P 0.0004). Pre-to-post nonlinear registration applied to FLAIR regions appeared to provide the best stratification between short- and long-term PFS and OS, as well as had the highest hazard ratio; however, statistical comparison of hazard ratios suggests that only linear fDMs applied to contrast-enhancing regions had a significantly different hazard ratio for both PFS and OS compared to other biomarkers. Results further suggest pre-to-post nonlinear fDMs evaluated in FLAIR abnormal regions had a higher sensitivity and specificity to 3-month and 6-month PFS compared to other measures

Nonlinear fDMs in Recurrent GBM

FIG. 2. Fractional volumes of fDM-classified tissues within contrast-enhancing and FLAIR abnormal regions for linear fDMs, pre-to-post nonlinear fDMs, and post-to-pre nonlinear fDMs and different registration algorithms/schemes (FNIRT applied to ADC, FNIRT applied to precontrast T1-weighted images, IRTK applied to ADC, IRTK applied to precontrast T1-weighted images). a: Fractional volume of tissue with a significant decrease in ADC (blue voxels) within contrast-enhancing regions. b: Fractional volume of tissue with a significant increase in ADC (red voxels) within contrast-enhancing regions. c: Fractional volume of tissue with a significant decrease in ADC (blue voxels) within FLAIR abnormal regions. d: Fractional volume of tissue with a significant increase in ADC (red voxels) within FLAIR abnormal regions. *P < 0.05; **P < 0.01; ***P < 0.001 for Tukeys test for multiple comparisons.

(Table 1). Similarly, results also suggest that pre-to-post nonlinear fDMs had a higher sensitivity and specificity to 6-month and 12-month OS compared to other measures (Table 2). Together, these results indicate that preto-post nonlinear fDMs applied to FLAIR abnormal regions provides the best clinical predictability, sensitivity, and specificity to both TTP and OS compared to linear fDMs or post-to-pre nonlinear fDMs applied to FLAIR or contrast-enhancing regions. DISCUSSION Voxelwise analysis of imaging parameters is a novel approach for both quantifying and visualizing potentially

heterogeneous tumor response to various cancer treatments. By examining voxelwise changes in ADC, the fDM technique has shown promising sensitivity to both cytotoxic (7,8) and antiangiogenic therapies (22). Despite these initial results, no significant improvements to the fDM technique have been made since the application of fDMs in regions of FLAIR abnormality (compared to region of contrast enhancement, exclusively). Since the development of fDMs in 2005, a linear registration scheme has been used for aligning ADC maps within the same patient over time. This technique poses several challenges when geometric distortions arise in ADC maps, or when misalignment of ADC maps occurs due to mass effect from growing tumor. These limitations to

Ellingson et al.

FIG. 3. Survival analyses for linear and nonlinear fDMs. a: Effects of linear and nonlinear fDM analysis within contrast-enhancing regions on time-to-progression (TTP). b: Effects of linear and nonlinear fDM analysis within contrast-enhancing regions on overall survival (OS). c: Effects of linear and nonlinear fDM analyses in FLAIR abnormal regions on TTP. d: Effects of linear and nonlinear fDM analyses in FLAIR abnormal regions on OS. *P < 0.05; **P < 0.01; ***P < 0.001 for Log-rank statistical analyses.

linear fDMs have been identified for some time; however, it has not been clear how best to implement an alternative registration scheme. For example, should baseline ADC maps be nonlinearly registered to each subsequent ADC map during follow-up (the pre-to-post scheme), or should each follow-up ADC map be nonlinearly registered to the baseline ADC map (the post-to-pre scheme). Alternatively, one might propose registering all ADC maps to a template space for subsequent analysis. As a first attempt to improve the fDM technique, we have implemented a simple nonlinear, elastic registration scheme. Results from this study suggest that nonlinear registration of ADC maps significantly changes the calculated fDMs and measured parameters, and these changes in fDM parameters tend to improve clinical predictability of fDMs with respect to bevacizumab treatment. Although our results suggest that nonlinear registration improves the ability for fDMs to predict survival, it is still not clear whether voxels classified as significantly

increasing or decreasing in ADC accurately depict physical changes in tumor microstructure or whether this is an artifact of image registration. To address this concern, the authors of a previous study have proposed using a similarity-guided correspondence correction algorithm for improving alignment of ADC maps for fDM analysis (25). Results from this pilot study suggested that fDM analysis was more accurate after application of this correction algorithm; however, no validation with respect to tissue or patient outcomes was presented. Therefore, tissue validation of nonlinear fDM-classified voxels is still required. Comparison to Other Biomarkers This study demonstrates a significant improvement in biomarker technology for use in predicting response of GBM to antiangiogenic therapies. Many imaging biomarkers have been explored for use in assessing the response to antiangiogenic therapies (26), including

Table 1 Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV) for Linear and Nonlinear fDM Metrics With Respect to 3-Month and 6-Month Progression-Free Survival 3-month progression-free survival Median % volume Sensitivity (%) Specificity (%) PPV (%) NPV (%) Sensitivity (%) Specificity (%) fDMs fDMs fDMs fDMs fDMs fDMs fDMs fDMs PPV (%) 6-month progression-free survival NPV (%)

Nonlinear fDMs in Recurrent GBM

Contrast-enhancing regions

VADC < 0.4 mm2/ms

VADC > 0.4 mm2/ms

FLAIR abnormal regions

VADC < 0.4 mm2/ms

VADC > 0.4 mm2/ms

Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear

24.80% 16.03% 17.28% 6.73% 3.44% 2.09% 19.64% 11.38% 17.14% 8.67% 5.20% 2.54%

60.00 55.00 60.00 50.00 50.00 45.00 55.00 65.00 55.00 60.00 55.00 50.00

54.00 52.00 54.00 50.00 50.00 48.00 52.00 56.00 52.00 54.00 52.00 50.00

34.29 31.43 34.29 28.57 28.57 25.71 31.43 37.14 31.43 34.29 31.43 28.57

77.14 74.29 77.14 71.43 71.43 68.57 74.29 80.00 74.29 77.14 74.29 71.43

52.27 54.55 54.55 45.45 52.27 52.27 59.09 59.09 54.55 59.09 63.64 54.55

53.85 57.69 57.69 42.31 53.85 53.85 65.38 65.38 57.69 65.38 73.08 57.69

65.71 68.57 68.57 57.14 65.71 65.71 74.29 74.29 68.57 74.29 80.00 68.57

40.00 42.86 42.86 31.43 40.00 40.00 48.57 48.57 42.86 48.57 54.29 42.86

Table 2 Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV) for Linear and Nonlinear fDM Metrics With Respect to 6-Month and 12-Month Overall Survival 6-month Overall survival Median % volume fDMs fDMs fDMs fDMs fDMs fDMs fDMs fDMs Sensitivity (%) Specificity (%) PPV (%) NPV (%) 12-month Overall survival Sensitivity (%) Specificity (%) PPV (%) NPV (%)

Contrast-enhancing regions

VADC < 0.4 mm2/ms

VADC > 0.4 mm2/ms

FLAIR abnormal regions

VADC < 0.4 mm2/ms

VADC > 0.4 mm2/ms

Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear Linear fDMs Pre-to-post nonlinear Post-to-pre nonlinear

24.80% 16.03% 17.28% 6.73% 3.44% 2.09% 19.64% 11.38% 17.14 8.67% 5.20% 2.54%

44.44 50.00 38.89 50.00 66.67 61.11 50.00 55.56 44.44 61.11 61.11 55.56

48.08 50.00 46.15 50.00 55.77 53.85 50.00 51.92 48.08 53.85 53.85 51.92

22.86 25.71 20.00 25.71 34.29 31.43 25.71 28.57 22.86 31.43 31.43 28.57

71.43 74.29 68.57 74.29 82.86 80.00 74.29 77.14 71.43 80.00 80.00 77.14

52.17 56.52 58.70 47.83 52.17 45.65 54.35 58.70 58.70 54.35 58.70 47.83

54.17 62.50 66.67 45.83 54.17 41.67 58.33 66.67 66.67 58.33 66.67 45.83

68.57 74.29 77.14 62.86 68.57 60.00 71.43 77.14 77.14 71.43 77.14 62.86

37.14 42.86 45.71 31.43 37.14 28.57 40.00 45.71 45.71 40.00 45.71 31.43

Ellingson et al.

positron emission tomography tracers (27), perfusion MRI (28), and other diffusion MR analyses (29); however, this is the first study to demonstrate the ability to predict both TTP and OS in recurrent GBM patients treated with bevacizumab. Additionally, linear fDM analyses have been successfully performed on patients treated with bevacizumab (22) by focusing on temporal changes in fDM-classified tissue volumes over time. Unlike this previous study, this study focused on using the initial changes after treatment to predict patient outcomes. Consistent with previous results, however, we found that patients having a larger volume of tissue with a lower ADC after treatment were more likely to progress sooner and have a shorter survival than patients having a lower volume of tissue with significantly lower ADC after treatment. Technical Limitations and Considerations In the current retrospective study, we have chosen to use ADC maps calculated from diffusion MR data that were collected with only b 0 s/mm2 and b 1000 s/mm2. By performing ADC calculations on only these DWIs, we ensure that our techniques can be retroactively applied across multicenter clinical trials. Despite this potential advantage, the use of a standard, clinical diffusion MR sequence parameters with only two b values poses potential limitations as to the accuracy of ADC measurements. The National Cancer Institute has recommended that ADC calculations should be performed using three or more DWIs at different b values (0, >100, and >500 s/ mm2) for perfusion-insensitive estimates of ADC. Additionally, the use of only two b values for calculation of ADC maps in fDM analyses can be confounded by other pathologies that may have been present, including cerebral ischemia. Another potential limitation is the nonlinear algorithm used for image registration. Many nonlinear image registration algorithms have been proposed and compared (30), and each has advantages and disadvantages. We chose to use FNIRT as part of the FSL software package for ease of use in our fDM postprocessing pipeline, as well as previous studies suggesting adequate performance (30). CONCLUSIONS In summary, nonlinear fDMs appear to provide a significant improvement in clinical predictability, sensitivity, and specificity of fDMs compared to the traditional, linear fDM approach. Additionally, pre-to-post nonlinear fDMs applied in FLAIR abnormal regions may provide the best sensitivity and specificity for progressing disease and survival compared to post-to-pre nonlinear fDMs. ACKNOWLEDGMENTS This research was supported by Brain Tumor Funders Collaborative (to W.B.P.), Art of the Brain (to T.F.C.), Ziering Family Foundation in memory of Sigi Ziering (to T.F.C.), Singleton Family Foundation (to T.F.C.), and Clarence Klein Fund for Neuro-Oncology (to T.F.C.).

REFERENCES
1. CBTRUS. Primary brain tumors in the United States 20042006. Chicago, IL: Central Brain Tumor Registry of the United States; 2010. 2. Stupp R, Mason WP, van den Bent MJ, Weller M, Fisher B, Taphoorn MJ, Belanger K, Brandes AA, Morosi C, Bogdahn U, Curschmann J, Janzer RC, Ludwin SK, Gorlia T, Allgeier A, Lacombe D, Cairncross JG, Eisenhauer E, Mirimanoff RO. Radiotherapy plus concomitant and adjuvent temozolomide for glioblastoma. N Engl J Med 2005;352: 987996. 3. Ellingson BM, Laviolette PS, Rand SD, Malkin MG, Connelly JM, Mueller WM, Prost RW, Schmainda KM. Spatially quantifying microscopic tumor invasion and proliferation using a voxel-wise solution to a glioma growth model and serial diffusion MRI. Magn Reson Med 2011;65:11311143. 4. Sugahara T, Korogi Y, Kochi M, Ikushima I, Shigematu Y, Hirai T, Okudo T, Liang L, Ge Y, Komohara Y, Ushio Y, Takahashi M. Usefulness of diffusion-weighted MRI with echo-planar technique in the evaluation of cellularity in gliomas. J Magn Reson Imaging 1999;9: 5360. 5. Ellingson BM, Malkin MG, Rand SD, Connelly JM, Quinsey C, LaViolette PS, Bedekar DP, Schmainda KM. Validation of functional diffusion maps (fDMs) as a biomarker for human glioma cellularity. J Magn Reson Imaging 2010;31:538548. 6. Chenevert TL, Stegman LD, Taylor JM, Robertson PL, Greenberg HS, Rehemtulla A, Ross BD. Diffusion magnetic resonance imaging: an early surrogate marker for therapeutic efficacy in brain tumors. J Natl Cancer Inst 2000;92:20292036. 7. Hamstra DA, Chenevert TL, Moffat BA, Johnson TD, Meyer CR, Mukherji S, Quint DJ, Gebarski SS, Fan X, Tsien CI, Lawrence TS, Junck L, Rehemtulla A, Ross BD. Evaluation of the functional diffusion map as an early biomarker of time-to-progression and overall survival in high-grade glioma. Proc Natl Acad Sci USA 2005;102: 1675916764. 8. Moffat BA, Chenevert TL, Lawrence TS, Meyer CR, Johnson TD, Dong Q, Tsien CI, Mukherji S, Quint DJ, Gebarski SS, Robertson PL, Junck L, Rehemtulla A, Ross BD. Functional diffusion map: a noninvasive MRI biomarker for early stratification of clinical brain tumor response. Proc Natl Acad Sci USA 2005;102:55245529. 9. Moffat BA, Chenevert TL, Meyer CR, McKeever PE, Hall DE, Hoff BA, Johnson TD, Rehemtulla A, Ross BD. The functional diffusion map: an imaging biomarker for the early prediction of cancer treatment outcome. Neoplasia 2006;8:259267. 10. Hamstra DA, Galban CJ, Meyer CR, Johnson TD, Sundgren PC, Tsien CI, Lawrence TS, Junck L, Ross DJ, Rehemtulla A, Ross BD, Chenevert TL. Functional diffusion map as an early imaging biomarker for high-grade glioma: correlation with convensional radiologic response and overall survival. J Clin Oncol 2008;26:33873394. 11. Ellingson BM, Malkin MG, Rand SD, Bedekar DP, Schmainda KM. Functional diffusion maps applied to FLAIR abnormal areas are valuable for the clinical monitoring of recurrent brain tumors. In: Proceedings of the 17th Annual Meeting of ISMRM, Honolulu, Hawaii, USA, 2009. p 285. 12. Ellingson BM, Rand SD, Malkin MG, Schmainda KM. Utility of functional diffusion maps to monitor a patient diagnosed with gliomatosis cerebri. J Neurooncol 2010;97:419423. 13. Ellingson BM, Malkin MG, Rand SD, Hoyt A, Connelly J, Bedekar DP, Kurpad SN, Schmainda KM. Comparison of cytotoxic and antiangiogenic treatment responses using functional diffusion maps in FLAIR abnormal regions. In: Proceedings of the 17th Annual Meeting of ISMRM, Honolulu, Hawaii, USA, 2009. p 1010. 14. Reese TG, Heid O, Weisskoff RM, Wedeen VJ. Reduction of eddy-current-induced distortion in diffusion MRI using a twice-refocused spin echo. Magn Reson Med 2003;49:177182. 15. Cox RW, Jesmanowicz A. Real-time 3D image registration for functional MRI. Magn Reson Med 1999;42:10141018. 16. Hamstra DA, Galban CJ, Meyer CR, Johnson TD, Sundgren PC, Tsien CI, Lawrence TS, Junck L, Ross DJ, Rehemtulla A, Ross BD, Chenevert TL. Functional diffusion map as an early imaging biomarker for high-grade glioma: correlation with conventional radiologic response and overall survival. J Clin Oncol 2008;26:33873394. 17. Andersson JL, Smith SM, Jenkinson M. FNIRTFMRIBs non-linear image registration tool. Annual Meeting of the Organization Human Brain Mapping, Melborne, Australia; 2008. 18. Andersson JL, Jenkinson M, Smith SM. Non-linear optimisation, FMRIB technical report TR07JA1; 2007.

Nonlinear fDMs in Recurrent GBM

19. Andersson JL, Jenkinson M, Smith SM. Non-linear registration, aka spatial normalisation, FMRIB technical report TR07JA2; 2007. 20. Rueckert D, Sonoda LI, Hayes C, Hill DL, Leach MO, Hawkes DJ. Nonrigid registration using free-form deformations: application to breast MR images. IEEE Trans Med Imaging 1999;18:712721. 21. Denton ER, Sonoda LI, Rueckert D, Rankin SC, Hayes C, Leach MO, Hill DL, Hawkes DJ. Comparison and evaluation of rigid, affine, and nonrigid registration of breast MR images. J Comput Assist Tomogr 1999;23:800805. 22. Ellingson BM, Malkin MG, Rand SD, Laviolette PS, Connelly JM, Mueller WM, Schmainda KM. Volumetric analysis of functional diffusion maps is a predictive imaging biomarker for cytotoxic and antiangiogenic treatments in malignant gliomas. J Neurooncol 2011;102: 95103. 23. Macdonald DR, Cascino TL, Schold Jr SC, Cairncross JG. Response criteria for phase II studies of supratentorial malignant glioma. J Clin Oncol 1990;8:12771280. 24. Pope WB, Sayre J, Perlina A, Villablanca JP, Mischel PS, Cloughesy TF. MR imaging correlates of survival in patients with high-grade gliomas. Am J Neuroradiol 2005;26:24662474. 25. Hoisak JD, Koh E-S, Yu E, Kassner A, Laperriere NJ, Menard C, Jaffray DA. An image similarity-guided correspondence correction for voxel-wise analysis applied to MR imaging of glioblastoma multiforme acquired pre- and post-chemotherapy. In: Proceedings of the 18th Annual Meeting of ISMRM, Stockholm, Sweden, 2010. p 616.

9
26. Pope WB, Young JR, Ellingson BM. Advances in MRI assessment of gliomas and response to anti-VEGF therapy. Curr Neurol Neurosci Rep 2011;11:336343. 27. Chen W, Delaloye S, Silverman DH, Geist C, Czernin J, Sayre J, Satyamurthy N, Pope W, Lai A, Phelps ME, Cloughesy T. Predicting treatment response of malignant gliomas to bevacizumab and irinotecan by imaging proliferation with [18F] fluorothymidine positron emission tomography: a pilot study. J Clin Oncol 2007;25:47144721. 28. Sorensen AG, Batchelor TT, Zhang WT, Chen PJ, Yeo P, Wang M, Jennings D, Wen PY, Lahdenranta J, Ancukiewicz M, di Tomaso E, Duda DG, Jain RK. A vascular normalization index" as potential mechanistic biomarker to predict survival after a single dose of cediranib in recurrent glioblastoma patients. Cancer Res 2009;69: 52965300. 29. Pope WB, Kim HJ, Huo J, Alger J, Brown MS, Gjertson D, Sai V, Young JR, Tekchandani L, Cloughesy TF, Mischel PS, Lai A, Nghiemphu PL, Rahmanuddin S, Goldin J. Recurrent glioblastoma multiforme: ADC histogram analysis predicts response to bevacizumab treatment. Radiology 2009;252:182189. 30. Klein A, Andersson J, Ardekani BA, Ashburner J, Avants B, Chiang MC, Christensen GE, Collins DL, Gee J, Hellier P, Song JH, Jenkinson M, Lepage C, Rueckert D, Thompson P, Vercauteren T, Woods RP, Mann JJ, Parsey RV. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. Neuroimage 2009; 46:786802.