A total of 69 patients with confirmed gliomas that were previously enrolled in a study of MR perfusion imaging at our institution were enrolled in the current retrospective study. All patients gave informed written consent according to guidelines approved by the Institutional Review Board at our institution. lists the study population data for each specific hypothesis tested.
Clinical MRI scans included a spoiled gradient recalled (SPGR) anatomical scan, pre-contrast T1-weighted (T1) scan, post-contrast T1-weighted (T1+C) scan, and a fluid-attenuated inversion recovery scan (FLAIR) collected on a 1.5-T MR scanner (Signa Excite, CVi, or LX; GE Medical Systems, Milwaukee, WI). 3D SPGR images were acquired with echo-time (TE)/repetition time (TR) = 3.16 msec/8.39 msec, number of averages (NEX) = 2, slice thickness = 1.3 mm collected contiguously, flip angle = 10 degrees, field-of-view (FOV) = 240 mm, and a matrix size of 256 × 192 (zero-padded and interpolated to 256 × 256) resulting in a total of 123 to 128 images. Axial pre- and post-contrast T1-weighted images were acquired before and after administration of up to 10cc gadobenate dimeglumine (Multihance; Bracco Diagnostics Inc., Princeton, NJ) or up to 20cc of gadodiamine (Omniscan; GE Healthcare Systems) contrast agent with a fast spin-echo (FSE) sequence, TE/TR = 24.16 msec/666.7 msec, NEX = 1, slice thickness of 5 mm with 1.5 mm interslice gap, flip angle = 90 degrees, FOV = 240 mm, and a matrix size of 256 × 192 (zero-padded and interpolated to 256 × 256) resulting in a total of 22 to 24 images. Axial FLAIR images were collected with a FSE readout, inversion time (TI) = 2,200 msec, TE/TR = 125.2 msec/10,000 msec, NEX = 1, slice thickness of 5 mm with 1.5 mm interslice gap, flip angle = 90 degrees, FOV = 240 mm, and a matrix size of 256 × 224 (reconstructed images were zero-padded and interpolated to 256 × 256) resulting in a total of 22 to 24 images.
Diffusion weighted images (DWIs) were collected with TE/TR = 102.2 msec/8,000 msec, NEX = 1, slice thickness of 5 mm with 1.5 mm interslice gap, flip angle = 90 degrees, FOV = 240 mm, and a matrix size of 128 × 128 (reconstructed images were zero-padded and interpolated to 256 × 256) using either an EPI or PROPELLER readout. DWIs were acquired with b = 0 and 1,000 s/mm2, using all gradients applied equally (isotropic). After collecting the images the apparent diffusion coefficient (ADC) images were calculated from the b = 1,000 s/mm2 and b = 0 images.
Functional Diffusion Maps (fDMs)
All images for each patient were registered to their own baseline, post-surgical, pre-treatment, SPGR anatomical images using a mutual information algorithm and a 12-degree of freedom transformation using FSL (FMRIB, Oxford, UK; http://www.fmrib.ox.ac.uk/fsl/
). Fine registration (1-2 degrees & 1-2 voxels) was then performed using a Fourier transform-based, 6-degree of freedom, rigid body registration algorithm (14
) followed by visual inspection to ensure adequate alignment. After proper registration was verified, voxel-by-voxel subtraction was performed between ADC maps acquired at subsequent time points and the baseline, post-surgical, pre-treatment, ADC maps to create ΔADC images. Individual voxels were stratified into three categories based on the change in ADC relative to baseline ADC maps. (Optimal ΔADC thresholds were explored in Hypothesis 2
). Red voxels represented areas within the tumor where ADC increased beyond the ΔADC threshold (i.e. “hypocellular” voxels), blue voxels represented areas within the tumor where ADC decreased beyond the ΔADC threshold (i.e. “hypercellular” voxels), and green voxels represented no change in ADC beyond the ΔADC threshold. This process is summarized in .
Figure 1 Calculation of functional diffusion maps from sequential apparent diffusion coefficient (ADC) maps. For each post-baseline time point, ADC maps from the current day are subtracted from baseline ADC maps. Each voxel within an ADC difference image is stratified (more ...)
Use of the terms “hypercellularity” and “hypocellularity” instead of “decreasing ADC” and “increasing ADC” in the current study may be misleading, since many pathologies and clinical scenarios can alter ADC measurements. As such, the possibility of localized infection, subacute stroke, substantial gliosis, tissue swelling from seizure activity, and changes in steroid use must be considered during interpretation of fDMs. At our institution, for example, clinical translation of the fDM technique involves integration of interpretations from board certified neurologists and radiologists, as well as biophysicists, to rule out the possibility of confounding factors.
Hypothesis 1: Glioma cell density is inversely proportional to ADC measurement
To test the hypothesis that tissue cell density is inversely proportional to ADC within brain tumor tissue, seventeen patients with a variety of glioma grades (WHO grades II-IV) who previously underwent closed diagnostic stereotactic biopsy via StealthSystem™ surgical navigation were retrospectively examined (). Intra-operative computed tomography images and post-operative 3D anatomical MR images were used to spatially localize the biopsy regions (). The ADC measurements corresponding to the precise spatial regions of the biopsy samples were extracted from the pre-operative ADC maps, after they were first registered to post-operative anatomical MR images. The volume of tissue examined in ADC maps ranged from 0.25 – 3 mL (approximately 200 – 2500 voxels on SPGR and registered/interpolated ADC maps).
Figure 2 Correlation between spatially-matched ADC measurements and cell density from stereotactic biopsy samples. A) Post-operative, high-resolution 3D T1-weighted anatomical MR images showing the biopsy location in a single patient. B) Representative histological (more ...)
Hematoxylin and eosin (H&E)-stained slides of the biopsy specimens prepared by the Pathology Core at our institution were used for cell density measurements. The slides from biopsy samples were analyzed using MetaMorph™image analysis software. The numbers of nuclei were manually counted by a single investigator, blinded to the diffusion imaging results, in two to four different regions on the histological slides at 20× magnification to provide a more random sampling of the cell density within the tumor specimen. Each slide was calibrated to physical units prior to cell counting to provide estimates of the number of nuclei per mm2, as well as the number of nuclei per high power field (HPF).
Linear regression was performed between the mean ADC measurement spatially matched to the biopsy site and the mean cell density obtained from the biopsy specimen for each patient. We assumed a linear model of the form:
where ADCTumor is the mean ADC measurement from the biopsy site (in mm2/s), B is the sensitivity of ADC measurement to cell density in units of ([mm2/s] / [nuclei/mm2]) or ([mm2/s]/[nuclei/HPF]), ρC is cell density (in nuclei/mm2 or nuclei/HPF), and C is the intercept.
Hypothesis 2: A measure of normal ADC variability across scan days must be determined in order to properly set the thresholds for fDMs
To estimate the reproducibity of voxel-wise ADC measurements the 95% C.I.s for different mixtures of NAWM, NAGM, and CSF were determined. Specifically ADC difference images (ΔADC images) created by subtracting a baseline ADC map from an ADC map were obtained at various post-baseline time points after image registration (see Functional Diffusion Maps section above for image registration details). The ADC maps were obtained at 1 week (n = 3 patients), 2 weeks (n = 3 patients), 1 month (n = 7 patients), 2 months (n = 10 patients), 3 months (n = 10 patients), 6 months (n = 10 patients), 9 months (n = 12 patients), and 1 year (n = 14) after baseline ADC maps to ensure the C.I.s take into consideration the possible range of ADC changes that might occur over the wide range of survival times observed in low to high grade gliomas. Patient information used for testing this hypothesis is summarized in .
Approximately 3,000 voxels from ΔADC images were obtained from each patient's NAWM, 3,000 voxels from each patient's NAGM, and 2,000 voxels from each patient's CSF within the subarachnoid space at each time point based on high resolution, 3D T1-weighted anatomical images () co-registered with ΔADC maps. If the data had equal variance according to Bartlett's test for equal variance, all voxels obtained from all patients for a particular (post-baseline) time point (e.g. at 1 month post-baseline) were pooled into a single distribution. Finally, all voxels from all patients and all time points were pooled to provide an overall distribution for calculation of ΔADC C.I.s for each tissue type. Note that the final distributions for NAWM+NAGM and NAWM+NAGM+CSF include the variability across multiple patients, tumor grades, post-baseline time points, tissue types (i.e. WM, GM, and CSF). Therefore, unavoidable artifactual covariance between tissue types may have occurred due to either mis-registration or partial volume averaging, along with actual variability that may have occurred due to global effects of standard therapy. Pooling the data by this method, therefore, allowed for a realistic estimate of the C.I.s that included many common factors found in our patient population.
Figure 3 Distribution of ΔADC for different tissue types and different post-baseline times. A) ΔADC histograms for normal-appearing white matter (NAWM). B) ΔADC histograms for normal-appearing gray matter (NAGM). C) ΔADC histograms (more ...)
Hypothesis 3: fDMs created with different ΔADC thresholds reflect different sensitivity and specificity to brain tumor progression
Five randomly chosen patients with recurrent glioblastoma having statistically similar total abnormal FLAIR volumes, from the 69 total patients enrolled in the current study, were used to test whether the 95% C.I.s for ΔADC (determined in Hypothesis 2
) produced physically different fDMs to those produced using the standard ΔADC threshold of 0.55 × 10-3
/s. Patient information used to test this hypothesis are summarized in . To test this hypothesis, the physical volumes of “hypercellularity” (voxels with decreased ADC relative to baseline) and “hypocellularity” (voxels with an increase in ADC relative to baseline) were calculated for these five patients using five ΔADC thresholds: 1) 95% C.I.s for NAWM, 2) 95% C.I.s for NAGM, 3) 95% C.I.s for a mixture of NAWM+NAGM, 4) 95% C.I.s for a mixture of NAWM+NAGM+CSF, and 5) the standard ADC thresholds recommended for human fDMs (10
). It is important to note that these thresholds were determined initially from different tissue types; however, these thresholds were then applied globally to all tissue types in these fDM patients (i.e. The 95% C.I. determined from NAWM, for example, was applied as the ΔADC threshold in fDM analysis in regions that may contain gray matter or white matter). These volumes were calculated exclusively for regions of the brain containing abnormal FLAIR signal intensity, which, when present, also contained all regions of gadolinium contrast enhancement. Although most fDM studies have focused only on contrast-enhanced regions, recent data have suggested that evaluating abnormal FLAIR regions are also clinically valuable (15
). A one-way, repeated measures ANOVA and Tukey's test for pairwise multiple comparisons were used to test this hypothesis.
In order to determine if different ΔADC thresholds used in fDMs applied to FLAIR regions differ in their sensitivity and specificity to progressive disease we closely examined 33 of the 69 patients enrolled in the current study who eventually showed disease progression according to our criteria for disease progression. Each patient's neurological and radiographic status during his or her clinical time-course was examined, and each session for each patient was categorized as either stable disease (SD) or progressing disease (PD). In this way, a single patient with many scan sessions may have sessions that were stable and a session that shows radiographic or neurological disease progression. “Progressive disease” was defined as having either neurological decline or radiographic progression, whereas “stable disease” was defined for all examinations that showed no change (or improvement) in neurological status and radiographic presentation.
To identify neurological progression we examined changes in the Karnofsky Performance Score (KPS). (The KPS scale ranges from 100 to 0, where 100 represents normal function and 0 represents expiration of the patient (17
). This scale is accepted as a valuable tool for examining the general neurological status of a patient (18
)). Fellowship-trained neuro-oncologists and trained physicians recorded KPS scores for each of the patients during routine clinical examinations, which were approximately every 8 weeks from the start of treatment unless shorter intervals were warranted. Neurological decline was defined on the basis of a greater than 20 point decrease in KPS score with respect to the previous exam, whereas a stable neurological exam was defined as no change or improvement in KPS score with respect to the previous exam.
Radiological status on a particular scan day was validated by a board certified radiologist by differential comparison with previous MR images (typically the images acquired approximately 8 weeks prior) using the Macdonald criteria when appropriate (19
). Briefly, radiological recurrence/progression was defined as new or enlarging regions of contrast-enhancement while the patients were on stable or increasing doses of corticosteroids during standard therapy (cytotoxic drugs and radiotherapy following surgical biopsy or resection).
Receiver-operator characteristic (ROC) analysis was used to determine the sensitivity and specificity of the rate of change in hypercellular volume (in uL/day) for different ΔADC thresholds to progressive disease. Note that we defined the rate of change in hypercellular volume as the change in hypercellular volume between two sequential fDMs, divided by the time interval between maps. We hypothesize that the rate of change in hypercellular volume is more sensitive, but less specific, to disease progression using lower ΔADC thresholds. A one-way analysis of variance was used to test whether the area under the ROC curves differed between ΔADC thresholds.