Comparison of tumor histology to dynamic contrast enhanced magnetic resonance imaging-based physiological estimates

Available online at www.sciencedirect.com

Magnetic Resonance Imaging 26 (2008) 1279 – 1293

Comparison of tumor histology to dynamic contrast enhanced magnetic resonance imaging-based physiological estimates Michael Aref a , Amir R. Chaudhari b , Keith L. Bailey c , Susanne Aref d , Erik C. Wiener e,f,g,⁎ a

Department of Nuclear, Plasma and Radiological Engineering, Beckman Institute Biomedical Imaging Center, College of Medicine, University of Illinois at Urbana-Champaign, Urbana, IL, USA b A.T. Still University of Health Sciences, Kirkoville, MO, USA c Drug Safety Research and Development, Pfizer Global Research and Development, Chesterfield, MO, USA d Aref Consulting Group LLC, Deland, IL, USA e Department of Radiology, University of Pittsburgh, Pittsburgh, PA, USA f Department of Biological Engineering, University of Pittsburgh Cancer Institute, Pittsburgh, PA, USA g Hillman Cancer Center, Pittsburgh, PA, USA Received 11 December 2007; accepted 21 February 2008

Abstract Purpose: The purpose of this study was to compare histologically determined cellularity and extracellular space to dynamic contrastenhanced magnetic resonance imaging (DCE MRI)-based maps of a two-compartment model's parameters describing tumor contrast agent extravasation, specifically tumor extravascular extracellular space (EES) volume fraction (ve), tumor plasma volume fraction (vp) and volume-normalized contrast agent transfer rate between tumor plasma and interstitium (KTRANS/VT). Materials and Methods: Obtained ve, vp and KTRANS/VT maps were estimated from gadolinium diethylenetriamine penta-acetic acid DCE T1-weighted gradient-echo images at resolutions of 469, 938 and 2500 μm. These parameter maps were compared at each resolution to histologically determined tumor type, and the high-resolution 469-μm maps were compared with automated cell counting using Otsu's method and a color-thresholding method for estimated intracellular (Vintracellular) and extracellular (Vextracellular) space fractions. Results: The top five KTRANS/VT values obtained from each tumor at 469 and 938 μm resolutions are significantly different from those obtained at 2500 μm (Pb.0001) and from one another (P=.0014). Using these top five KTRANS/VT values and the corresponding tumor EES volume fractions ve, we can statistically differentiate invasive ductal carcinomas from noninvasive papillary carcinomas for the 469- and 938-μm resolutions (P=.0017 and P=.0047, respectively), but not for the 2500-μm resolution (P=.9008). The color-thresholding method demonstrated that ve measured by DCE MRI is statistically similar to histologically determined EES. The Vextracellular obtained from the color-thresholding method was statistically similar to the ve measured with DCE MRI for the top 10 KTRANS/VT values (PN.05). DCE MRIbased KTRANS/VT estimates are not statistically correlated with histologically determined cellularity. Conclusion: DCE MRI estimates of tumor physiology are a limited representation of tumor histological features. Extracellular spaces measured by both DCE MRI and microscopic analysis are statistically similar. Tumor typing by DCE MRI is spatial resolution dependent, as lower resolutions average out contributions to voxel-based estimates of KTRANS/VT. Thus, an appropriate resolution window is essential for DCE MRI tumor diagnosis. Within this resolution window, the top KTRANS/VT values with corresponding ve are diagnostic for the tumor types analyzed in this study. © 2008 Published by Elsevier Inc. Keywords: Cellularity; Extravascular extracellular space; EES; Dynamic contrast-enhanced; DCE; Magnetic resonance imaging; MRI; Two-compartment model; Tumor; Contrast agent extravasation; Volume fraction; Contrast agent transfer rate; KTRANS

1. Introduction

⁎ Corresponding author. Tel.: +1 412 623 4658; fax: +1 412 623 3355. E-mail address: [email protected] (E.C. Wiener). 0730-725X/$ – see front matter © 2008 Published by Elsevier Inc. doi:10.1016/j.mri.2008.02.015

Dynamic contrast-enhanced magnetic resonance imaging (DCE MRI) has become an important tool in diagnosing breast carcinoma, in differentiating between benign and malignant breast lesions, and in monitoring antiangiogenic

1280

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

therapy [1]. This technique has been shown to have a high sensitivity (83–100%) and a wide range of specificity (30– 100%), which may be attributed to the significant heterogeneity of breast lesions [1–4]. Introduction of extracellular gadolinium-based contrast agents results in improved detection of malignant tumors [3,5]. A bolus of a lowmolecular-weight contrast agent can be tracked as it reaches the capillary bed and diffuses across capillaries into the extravascular extracellular space (EES) of the tumor at a rate determined by blood flow, microvessel permeability and surface area [1,4,6,7]. Approximately 12–45% of lowmolecular-weight contrast media leaks into the EES during the first pass but does not cross cell membranes [7]. The total uptake of contrast agent in the interstitial space is related to maximum enhancement, and the clearance of contrast agent is associated with the washout rate [1]. Tumors with high vascularity exhibit early signal enhancement with rapid washout of intravenous contrast, while slow and sustained contrast enhancement is correlated with low tumor vascularity [2]. Most malignant and benign lesions exhibit different washout patterns on DCE MRI [2–4]. Malignant lesions typically show a faster and stronger signal intensity rise after injection of gadolinium-based contrast agents, while benign tumors exhibit either persistent or plateau-type enhancement [3,4]. Angiogenesis is directly associated with aggressive tumor growth and inversely associated with disease-free survival, and both are measured by intratumoral microvessel density (iMVD) of invasive breast carcinoma [1,8–10]. Tumor synthesis and secretion of vascular endothelial growth factor (VEGF), which is also known as vascular permeability factor, increase the permeability of capillaries and are thought to lead to angiogenesis and stroma generation [11]. Oncogenes such as ras, raf and src are associated with overexpression of VEGF, which plays a pivotal role in the development of new tumor blood vessels, increased vessel density, proliferation of endothelial cells and increased microvessel permeability [12–16]. High vessel densities in tumors often result in aggressive tumor growth and poor prognosis [8–10,16,17]. Contrast-enhanced MRI that is correlated with traditional histopathology and tumor angiogenesis measured by iMVD in human tumors shows significant increases in iMVD with increasing histological tumor grade. Different microvessel densities can be found within different regions of the same tumor, and areas with high microvessel density are referred to as “hot spots” [2]. One in vitro histological technique used for differentiating benign from malignant tumors relies on differences in capillary density “hot spots” resulting from heterogeneous VEGF-induced angiogenesis or neovascularization [10,18–21]. Weidner et al. related prognostic outcome to the number of vessels counted in a “hot spot” using a microscope field of view (FOV) of 0.740 mm2 (860 μm diameter). A microscope FOV of 0.152 mm2 (390 μm diameter) was the greatest magnification at which capillary densities significantly distinguished benign from

malignant tumors [20,21]. Any area between these two limits can be used to grade and thus accurately distinguish benign from malignant tumors. Vascular “hot spots” are mostly found at the edges of invasive tumors [22]. Early rim enhancement is observed in DCE MRI using gadopentetate dimeglumine as a contrast agent in breast lesions with high ratios of peripheral to central microvessel densities and VEGF expression [15]. This result suggests that VEGF induces angiogenesis and increases extravasation of gadopentetate dimeglumine from intravascular to interstitial tissues in breast tumors and that peripheral enhancement is a strong indicator of malignancy [1,15]. We test the hypothesis that the low spatial resolution used in clinical DCE MRI results in partial volume effects that yield inaccurate physiological parameters describing tumor contrast agent extravasation, specifically tumor EES volume fraction (ve), tumor plasma volume fraction (vp) and volume-normalized contrast agent transfer rate between tumor plasma and interstitium (KTRANS/VT), resulting in erroneous diagnostic information when compared with histologically determined cellularity and EES. We used DCE T1-weighted gradient-echo multislice images to measure the KTRANS/VT value obtained at a standard clinical resolution (2500×2500 μm in-plane resolution) compared to that obtained with higher-resolution techniques (938×938 and 469×469 μm in-plane resolution) for gadolinium diethylenetriamine penta-acetic acid (GdDTPA). We compared the top 10 values of estimated intracellular (Vintracellular) and extracellular (Vextracellular) spaces for each high-resolution digital image tumor histological section, with KTRANS/VT and ve measured with DCE MRI at the top 10 values of KTRANS/VT for each MR image at 469 μm resolution. Our results show that the correlation between tumor KTRANS/VT and tumor type and grade is spatial resolution dependent and can be diagnostically differentiated at 469 and 938 μm in-plane resolution, but not at clinical spatial resolution (2500 μm). Furthermore, at the same microscope and imaging resolution of 469 μm, histologically estimated Vextracellular is statistically the same as DCE MRI-estimated ve. 2. Materials and methods 2.1. Animal model and preparation The tumor model used was a female Sprague–Dawley rat with chemically induced mammary tumors. This model produces tumors that are pathologically and histologically very similar to breast tumors in humans. Tumors were induced at 30 days of age with an intraperitoneal injection of 180 mg of N-ethyl-N-nitrosourea in 10% DMSO per kilogram of rat [23]. The tumors developed over a period of 2–6 months, and the animals were imaged when the tumors were ≥1 cm in diameter. Prior to imaging, animals were anesthetized with 68 mg of ketamine plus 14 mg of xylazine and 2.3 mg of acepromazine per kilogram body weight via intramuscular injection to the lateral thigh. The tail vein was catheterized with a 0.41-ml

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

25-gauge 0.75-in. butterfly catheter for intravenous access, and the animal was subsequently injected with 0.5 ml of 2 U/100 ml heparinized saline. The animal was placed in a custom-made bed designed for the imaging coil used along with a 1-mM Gd-DTPA (Magnevist; Berlex Laboratories) imaging fiduciary next to the right flank. 2.2. MRI MR images with spatial resolutions of 469 μm [(read out [RO] 24 cm/512)×(phase encode [PE] 6 cm/128)] were acquired using a SISCO 4.7-T/33-cm bore system. Both scout imaging and dynamic imaging used a fast T1-weighted gradient-echo multislice pulse sequence (TR=70 ms; TE=4.7 ms; flip angle=80°; slice=7; thickness=2 mm; coronal orientation, FOV=(RO 24 cm/512)×(PE 6 cm/128), average=4. Each 35-s dynamic acquisition was separated from the next by a 10-s delay. A dose of 0.3 mmol kg−1 Gd-DTPA was introduced into the bloodstream following the first acquisition. The data were read into MATLAB (The Mathworks, Inc., Natick, MA) using an algorithm written by Dr. Jeff Tsao. 2.3. Subset image data preparation Using the pulse sequence and image data described above, the image resolution was arrayed between 2500 μm (clinical) and 469 μm (obtained). That is, lower-resolution images were formed from central k-space subsets of the high-resolution (469 μm) images [24,25]. The following resolutions were analyzed:

The relative signal intensity obtained from the reprocessed images was calculated from the signal intensity of a region(s) of interest (ROI; e.g., a single voxel within the tumor or a ventricle of the heart) over the signal intensity of the image fiduciary. Relative signal intensities were converted to contrast agent concentration by a calibration curve: 0 1

½CAc

C C1 C RI þ RI0  RIPRE A 1 C0 C2

The dosage D (mmol kg−1) is the known administered dose, a1 and a2 (kg L−1) are the concentration amplitudes, α (min−1) is the distribution rate constant and β (min−1) is the excretion rate constant. The plasma curve values a1,2, α and β were fitted for by a nonlinear least squares fitting by the Gauss–Newton method [26] and used as constants in the fit of the tumor compartment equation: 0 0 1 B B B ½CAt ðt ÞcDB @ a1 @ v p þ 0 B þa2 B @vp þ

B B @

2.4. Image analysis

B lnB @

That is, the standard curve is zeroed to the observed preinjection RI. For the plasma compartment with GdDTPA at 4.7 T and 37°C, C0=3.39 mM−1, C1=0.951 mM−1 and C2=0.259 mM−1. For the tumor compartment with GdDTPA at 4.7 T and 37°C, C0=1.526 mM−1, C1=0.8583 mM−1 and C2=1.276 mM−1. To account for differences between animals, individual pharmacokinetic parameters were measured by converting relative intensities to contrast agent concentrations from image slices through the heart, which were then fitted by a two-compartment model's plasma compartment equation:     ð2Þ CAp ðt Þ ¼ D a1 eat þ a2 ebt

0

• (RO 24 cm/96)×(PE 6 cm/24) [2500 μm] • (RO 24 cm/256)×(PE 6 cm/64) [938 μm] • (RO 24 cm/512)×(PE 6 cm/128) [469 μm].

ð1Þ

where RI is the relative intensity of the tumor, RI0 (=C0(1− C1)) is the theoretical relative intensity in the absence of the contrast agent (i.e., the theoretical preinjection relative intensity of the tumor or plasma), RIPRE is the actual preinjection relative intensity of the tumor or plasma, [CA] is the contrast agent concentration and C0,1,2 are the measured or calculated coefficients for converting RI to [CA]. The (RI0−RIPRE) term serves to scale the conversion between RI and [CA] due to RI differences between tumor phantoms and the actual tumor, or between theoretical and actual plasma RI.

1281

C at ve Ce Vt a A 1 KTRANS 1

C bt ve Ce Vt b A 1 KTRANS

1

1

C KTRANS t C a1 v e a2 v e Ce Vt C þ A Vt a Vt b A 1 1 KTRANS KTRANS

ð3Þ

where the fitting parameters are ve (=Vt/VT)=the tumor EES volume fraction, vp=the tumor plasma volume and KTRANS/ VT (min−1 )=the tumor-volume-normalized transfer rate between plasma and tumor EES. Vt (L kg−1) is the tumor EES volume filled by a specific contrast agent, and VT (L kg−1) is the inverse tumor density. The parameters ve, vp and KTRANS/VT were fitted by a multiple β0 (initial guess) nonlinear least squares fitting by the Gauss–Newton method [26] on a voxel-by-voxel basis. The whole tumor ROI was fitted; from those parameters, an array of β0 for each voxel was determined. Each mapped point has an F test for P values and r2 [27]. The initial condition (preinjection) tumor contrast agent concentration was forced; that is, a [CAt(t=0)]=0 data point was added. For each voxel, only the best fit was used, and these mapped points were further filtered: points (a) that did not converge; (b) whose initial t=0 point did not start less than the maximum fitted tumor contrast agent concentration; (3) that were physiologically unrealistic (i.e., the fitted values must be 0≤veb1, 0≤vpb1 and 0≤KTRANS/VT); or (4) that fitted poorly (r2≤.5) are set to zero. All data analyses were performed with MATLAB (The Mathworks, Inc.).

1282

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293 Table 2 Analysis number and the corresponding histopathologically determined tumor type Animal

Tumor type

1a 1b

IDC Noninvasive papillary carcinoma+ ductal adenoma+tubular adenoma IDC Noninvasive papillary carcinoma IDC Fibroma Early fibroadenoma Fibroadenoma IDC Early IDC Noninvasive papillary carcinoma

2 3a 3b 4a 4b 5 6a 6b 7

Fig. 1. The plasma curve for Gd-DTPA.

2.5. Histopathology Following the imaging studies, the rats were sacrificed using 2 ml of sodium pentobarbital (Nembutal) intraperitoneally and transcardially perfused with 0.1 M phosphate buffer. The tumors were centrally cross-sectioned based on scout images, which maintained the orientation planes of the tumor in situ. The tissue slices were sectioned at a thickness of 3 μm and stained with hematoxylin–eosin for routine histopathology. Classification of the mammary tumors was in accordance with the World Health Organization's International Agency for Research on Cancer [28]. 2.6. Tissue section selection criteria From the histopathology, 10 stained tissue sections were selected for high-resolution digital image acquisition. Tumor selection criteria were based on non-motion-compromised areas (i.e., areas without motion artifacts due to breathing, heartbeat or animal motor actions). 2.7. Microscope/image acquisition High-resolution digital images of the tumor slices were created with the University of Illinois at Urbana-Champaign

Beckman Institute Microscopy Suite's Zeiss Axiovert 100 inverted research-grade fluorescence light microscope using a ×20 Zeiss plan neofluar objective with a numerical aperture of 0.50, an infinite tube length and a working distance of 0.17 mm. Images were captured with a Photometrics CoolSNAP fx cooled color CCD camera with a camera resolution of 1300×1028 pixels and a pixel size of 0.338 μm pixel−1. The MCID Elite 7.0 image tiling software was used to create a series of image tiles that were aligned and stitched in the Visualization, Media and Imaging Laboratory, Beckman Institute, University of Illinois at Urbana-Champaign, using custom software written by the Imaging Technology Group. 2.8. Histological image data analysis The high-resolution digital images were read into MATLAB (The Mathworks, Inc.) to perform cell counting using Otsu's method and color-thresholding method. The image-processing algorithm using Otsu's method converted user-selected color images of cellular, extracellular (stromal) and ductal objects to gray-scale images before graylevel thresholding was applied and cell counting was performed. Image analysis by color thresholding was performed to separate objects based on the intensity of the colors red, green and blue on the digital images before cell counting was performed. The top 10 estimates of intravascular and extravascular spaces for each FOV obtained from each histological slice were chosen and

Table 1 Whole-body pharmacokinetic values for tumor-bearing rats Animal

Ve (L kg−1) Vp (L kg−1) Kp↔e (L kg−1 min−1) Kp→k (L kg−1 min−1) a1 (kg L−1) a2 (kg L−1) α (min−1) β (min−1) TD (min) TE (min)

1 2 3 4 5 6 7 Average S.D.

0.0141 0.0295 0.0330 0.0336 0.0400 0.0316 0.0277 0.030 0.008

0.0337 0.0356 0.0362 0.0340 0.0351 0.0355 0.0357 0.0351 0.0009

0.00327 0.00282 0.00382 0.00636 0.00374 0.00819 0.00357 0.005 0.002

0.00142 0.00209 0.00143 0.00116 0.00230 0.00213 0.00125 0.0017 0.0005

8.70 12.7 13.1 14.6 15.1 13.3 12.2 13 2

20.9 15.3 14.4 14.8 13.3 14.9 15.8 16 2

0.328 0.175 0.221 0.376 0.200 0.490 0.229 0.3 0.1

0.0297 0.0321 0.0206 0.0172 0.0307 0.0317 0.0197 0.026 0.006

2.11 3.96 3.13 1.84 3.46 1.42 3.03 2.7 0.9

23.3 21.6 33.6 40.4 22.6 21.9 35.2 28 8

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

1283

Fig. 2. (A) MR image ROI at 469 μm resolution. (B) A high-resolution image was acquired from the histology slide. A much reduced version of this image is shown here (71.5×81.8 μm).

compared with the histologically determined cellularity and capillary density on DCE MRI-based maps of the twocompartment model's parameters, namely, tumor EES volume fraction (ve), tumor plasma volume fraction (vp) and volume-normalized contrast agent transfer rate between tumor plasma and EES (KTRANS/VT). Once high-resolution digital images of tumor sections had been read into MATLAB, the ROI was selected manually by the user, which was segmented into 469×469 μm2 sections. Image segmentation started from the upper left corner of the ROI and continued across, extracting 469×469 μm2 sections until it had reached the upper right corner of the ROI. Image segmentation proceeded with the next row until the entire ROI had been completely segmented. Single nuclei, groups

of nuclei, fibrous tissue and duct space were the objects manually selected by the user from a 469×469 μm2 image section. These objects were used to calculate color threshold limits on cellular, stromal and ductal spaces. Implementing Otsu's method required the conversion of the color image to gray-scale format using the im2bw function in MATLAB. The graythresh function in MATLAB was applied to compute a global threshold that was used to convert the gray-scale image to a binary image. The intensity value of each pixel was compared with the upper and lower threshold limits, and pixels that fell within the threshold limits were given a value of 1; otherwise, it was set to 0. A connectivity parameter was used to compare each pixel with either four or eight neighboring pixels. The

Fig. 3. Histology, Areaintracellular and Areaextracellular of a 469×469 μm2 section using color-thresholding method, with the user selecting one nucleus.

1284

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

from the areas within the color threshold of the selected cellular objects minus the areas of cellular threshold limits overlapping with fibrous or ductal space objects. Vextracellular was estimated from the areas within the color threshold of the selected fibrous objects minus the areas of fibrous threshold limits overlapping with ductal space objects. The top 10 estimated Vintracellular and Vextracellular values of each image were compared with top 10 KTRANS/VT values and corresponding ve values, respectively, obtained from DCE MRI tumor extravasation parameter maps for each tumor. 2.9. Statistical analysis

Fig. 4. Map of estimated Vintracellular and Vextracellular using color-thresholding method, with the user selecting one nucleus.

For each tumor and fast Fourier transform (FFT) spatial resolution, the means of the top five KTRANS/VT values were used in a randomized complete-block-design analysis of variance (ANOVA). The least squares means of log KTRANS/ VT for the 469, 938 and 2500 μm resolutions were compared. For each tumor at 469 μm resolution, the means of the top five KTRANS/VT values were used in a randomized completeblock-design ANOVA. For the tumor types and the 469 μm resolution FFT, the means of the top five KTRANS/VT values and corresponding ve and vp values were used in a nesteddesign ANOVA. Tumor types were compared based on (a) the least squares means of log KTRANS/VT; (b) the least squares means of log KTRANS/VT and ve; and (c) the least pffiffiffiffiffi squares means of log KTRANS/VT and vp .

clustered pixels with a value of 1 in the binary image were labeled as an object. The estimated intracellular space (Vintracellular) was calculated from labeled cellular objects minus areas of overlap between labeled cellular objects with labeled ductal space objects. The estimated extracellular space (Vextracellular) was calculated from labeled fibrous objects minus areas of overlap between labeled fibrous objects with labeled ductal space objects. The top 10 estimated Vintracellular and Vextracellular values of each image section were compared with top 10 KTRANS/VT values and corresponding ve values, respectively, obtained from DCEMRI tumor extravasation parameter maps for each tumor for both connectivity parameters. The color-thresholding method was applied to the same separate objects described for Otsu's method based on their mean color intensities. The objects manually selected by the user produced distinct peaks, which were used to define the boundaries of each object when separated into its red, green and blue components. Each peak was labeled, and the upper and lower threshold limits for each peak were set at 1 S.D. around the mean color intensity. Each pixel in the color image was compared with the threshold limits of each peak, and clusters of pixels that fell within the set threshold limit of a peak were counted as an object. Vintracellular was estimated

The analysis of the plasma compartment averaged over all seven animals: Ve=0.03±0.008 L kg−1, Vp=.0351±0.0009 L kg−1, Kpe=0.005±0.002 L kg−1 min−1, Kpk=0.0017±0.0005 L kg−1 min−1, TD=2.7±0.09 min and TE=28±8 min. These physiological parameters can be used to find the twocompartment model's parameters that are needed in Eq. (2): a1=13±2 kg L−1, a2=16±2 kg L−1, α=0.3±0.1 min−1 and β=0.026±0.006 min−1 (Fig. 1 and Table 1). For each tumor analyzed, the specific whole-body physiological parameters of the animal bearing the tumor were used. The whole-body physiological parameters obtained are in good agreement with literature values. The average blood plasma volume of a rat Vp=0.0350 L kg−1 [29], while the excretion half-life TE falls within values obtained

Table 3 Comparison between the ve measured with DCE-MRI at KTRANS/VT “hot spots” and the Vextracellular measured with Otsu's method and colorthresholding method

Table 4 Comparison between the KTRANS/VT “hot spots” measured with DCE-MRI and the Vintracellular measured with Otsu's method and color-thresholding method

Histology

Cell method

P

Histology

Cell method

P

Color-thresholding method Color-thresholding method Otsu's method Otsu's method

Groups One nucleus Groups One nucleus

.987 .928 .0177 .0376

Color-thresholding method Color-thresholding method Otsu's method Otsu's method

Groups One nucleus Groups One nucleus

.0009 b.0001 b.0001 b.0001

3. Results 3.1. Plasma compartment

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

1285

Fig. 5. Histology, Areaintracellular and Areaextracellular of a 469×469 μm2 section using color-thresholding method, with the user selecting a group of nuclei.

by other groups: Aref et al. [5] reported 25 min, Bieser [30] reported 23±7 min and Brasch [31] reported 20 min. In addition, the Gd-DTPA distribution half-life is shorter than its excretion half-life, since in order for the contrast agent to be detected in the tissues, it must be distributed before being excreted. In addition, KpeNKpk, in agreement with the assumptions involved in deriving the twocompartment model [5,32]. Of the nine animals used in this study, seven animals bearing 13 tumors were used. Two were discarded because their whole-body pharmacokinetic parameters (i.e., a1,2, α and β) were well outside the limits of normal. That is, animals with physiologically long excretion half-lives (TEN100 min) or large EES volume (VeN0.0500 L kg−1), most likely due to anesthesia overdose or intolerance, were dropped.

(Figs. 5 and 6). Ductal spaces were not misidentified, and some minimal artifacts appeared at the tiling interfaces. Maps of Vintracellular and Vextracellular agreed with gross inspection of microscope slides. Comparisons with the DCE MRI values show that ve and Vextracellular are statistically similar (P=.928), while KTRANS/VT and Vintracellular are statistically different (Pb.0009) (Tables 3 and 4). One-nucleus Otsu's method may qualitatively underestimate fibrous intracellular areas and nuclei extracellular areas (Figs. 7–9). Duct spaces are not misidentified, and some minimal artifact areas are caused at the tiling interfaces. Maps of cell counts based on connectivity parameters of four and eight agree with gross inspection of microscope slides; however, the connectivity parameter of eight may be a better

3.2. Histological analysis The 13 tumors were identified by histological analysis under a microscope, independent of KTRANS/VT mapping, as 5 invasive ductal carcinomas (IDC), 3 noninvasive papillary carcinomas, 2 fibroadenomas, 1 ductal adenoma, 1 tubular adenoma and 1 fibroma (Table 2 and Fig. 2). One nucleus color-thresholding method shows a qualitatively reasonable estimation of intracellular areas, with a mild underestimation of fibrous extracellular areas (Figs. 3 and 4). Ductal spaces are not misidentified. Some minimal artifacts areas are caused at the tiling interfaces. Maps of Vintracellular and Vextracellular agreed with gross inspection of microscope slides. When compared with the DCE MRI values, ve and Vextracellular are statistically similar (P=.928), while KTRANS/VT and Vintracellular are statistically different (Pb.0001) (Tables 3 and 4). Group of nuclei color-thresholding method shows better qualitative estimation of intracellular areas and fibrous extracellular areas than the one-nucleus thresholding method

Fig. 6. Map of estimated Vintracellular and Vextracellular of tumor using colorthresholding method, with the user selecting a group of nuclei.

1286

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

Fig. 7. Histology, Areaintracellular and Areaextracellular of a 469×469 μm2 section using Otsu's method, with the user selecting one nucleus.

representation of the actual cell count. Maps of Vintracellular and Vextracellular also agree with gross inspection of microscope slides. Comparisons between ve and Vextracellular and between KTRANS/VT and Vintracellular are statistically different and yielded P=.0376 and Pb.0001, respectively (Tables 3 and 4). Group-of-nuclei Otsu's method shows a qualitative underestimation of intracellular fibrous areas and overestimation of intracellular nuclei areas. Extracellular fibrous and nuclei areas show qualitatively reasonable estimations (Figs. 10–12). Maps of cell counts based on connectivity parameters of four and eight, and maps of Vintracellular and Vextracellular agree with gross inspection of microscope slides. Comparisons between ve and Vextracellular and between KTRANS/VT and Vintracellular are statistically different and yielded P=.0177 and Pb.0001, respectively (Tables 3 and 4).

The 469 μm resolution (0.220 mm2) and 938 μm resolution (0.880 mm2) KTRANS/VT maps have voxel sizes that compare with the clinically relevant microscope FOV range from 0.152 mm2 (390 μm diameter) to 0.740 mm2 (860 μm diameter) reported by Weidner [10,19–21] et al. The final image resolution analyzed (2500 μm) is that used in current clinical DCE MRI. Comparing the maps at the three resolutions analyzed shows that the volume-normalized contrast agent transfer rate between tumor plasma and interstitium (KTRANS/VT) heterogeneity begins to be averaged out as a function of resolution (Tables 5 and 6; Figs. 13–19). In addition, the 469 μm resolution data had KTRANS/VT values significantly greater than the two other resolutions. The

Fig. 8. Map of cell count based on connectivity parameters of four and eight using Otsu's method, with the user selecting one nucleus.

Fig. 9. Map of estimated Vintracellular and Vextracellular using Otsu's method, with the user selecting one nucleus.

3.3. Effects of imaging resolution

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

1287

Fig. 10. Histology, Areaintracellular and Areaextracellular of a 469×469 μm2 section using Otsu's method, with the user selecting a group of nuclei.

top five KTRANS/VT values obtained from the 469 μm resolution are statistically different from those obtained at the 938 μm resolution and the clinical imager resolution (2500 μm). The ANOVA of the top five KTRANS/VT values from the 469 μm resolution compared to the 938 and 2500 μm \resolutions yielded P=.0014 and Pb.0001, respectively. The ANOVA of the 938 μm resolution compared to the 2500 μm resolution of top five KTRANS/VT values yielded Pb.0001 (Table 7). Thus, the clinical imager resolution misses or averages out relevant contrast agent transfer coefficients that could be detected at higher resolutions (469 and 938 μm) near Weidner's range. Partial voluming does affect KTRANS/VT measurements. The decreased dynamic range due to the effects of an increased partial volume can be seen in compared maps. The quality of fit improves with lower resolution (Figs. 17–19).

Due to the distribution of tumors in this experiment, only the IDC (n=5), the noninfiltrating papillary carcinoma (n=3) and their respective fitting parameters were used to see whether a differential diagnosis is possible. Tumor types were compared based on (a) top five KTRANS/VT values; (b) top five KTRANS/VT values and the corresponding ve values; and (c) top five KTRANS/VT values and the corresponding vp values. On average, the IDC pharmacokinetic parameters based on FFT reconstruction at 469 μm resolution were ve=0.035±

Fig. 11. Map of cell count based on connectivity parameters of four and eight using Otsu's method, with the user selecting a group of nuclei.

Fig. 12. Map of estimated Vintracellular and Vextracellular using Otsu's method, with the user selecting a group of nuclei.

Increasing voxel size or decreasing spatial resolution averages out error due to noise from physiological motion or hardware instability, but also averages out diagnostically important physiological information. 3.4. DCE-MRI-based tumor diagnosis

1288

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

Table 5 Average and standard deviations of ve, vp and Kp↔t/VT as a function of resolution for IDC (n=5) 469 μm

Average S.D.

938 μm

2500 μm

ve

vp

Kp↔t/VT [L (kg min)−1]

ve

vp

Kp↔t/VT [L (kg min)−1]

ve

vp

Kp↔t/VT [L (kg min)−1]

0.035 0.009

0.001 0.001

0.03 0.01

0.031 0.007

0.0009 0.0007

0.016 0.007

0.04 0.01

0.002 0.001

0.008 0.005

0.009, vp=0.001±0.001 and KTRANS/VT=0.03±0.01 L kg−1 min−1. At 938 μm resolution, ve=0.031±0.007, vp=0.0009± 0.0007 and KTRANS/VT=0.016±0.007 L kg−1 min−1. Finally, at 2500 μm resolution, ve=0.04±0.01, vp=0.002±0.001 and KTRANS/VT=0.008±0.005 L kg−1 min−1 (Table 5). The noninfiltrating papillary carcinoma had average pharmacokinetic parameters based on FFT reconstruction for 469 μm resolution: ve=0.06±0.02, vp=0.002±0.003 and KTRANS/VT=0.04±0.03 L kg−1 min−1. At 938 μm resolution, ve=0.05±0.02, vp=0.004±0.004 and KTRANS/VT=0.03±0.02 L kg−1 min−1. Finally, at 2500 μm resolution: ve=0.03±0.02, vp=0.001±0.001 and KTRANS/VT=0.012±0.008 L kg−1 min−1 (Table 6). The 469 and 938 μm resolutions enabled us to differentiate between two tumor types based on the ve and KTRANS/VT. The first comparison, ANOVA of the top five KTRANS/VT values grouped by tumor type, could not statistically differentiate the IDC from the noninfiltrating papillary carcinomas. For 469, 938 and 2500 μm resolutions, P=.7555, P=.6565 and P=.6063, respectively. The corresponding vp and ve values were statistically tested and shown to be dependent on the top five KTRANS/VT values. When the corresponding vp values for the top five KTRANS/VT “hot spots” were incorporated into the ANOVA, the tumor types could still not be statistically differentiated. However, when the corresponding ve values for the top five KTRANS/VT values were incorporated into the ANOVA, the tumor types could be statistically differentiated for 469 and 938 μm resolutions (P=.0017 and P=.0047, respectively), but not for 2500 μm resolution (P=.9008) (Table 8).

of tumor angiogenesis regulate the extraction of an agent by a tumor from the blood. Contrast agent extraction by tumor tissue depends on (a) capillary surface area S (m2 kg−1); (b) capillary permeability P (m min−1); (c) capillary blood perfusion F (m3 (kg min)−1); (d) transit time of the agent through the tumor interstitium; and (e) the plasma half-life TD and TE [35–37]. Hypothetically, measurements of contrast agent extravasation can be correlated with tumor type if PS, F and interstitial transit time are histologically represented by cellular, capillary and stromal densities, as well as the presence of angiogenic factors such as VEGF. All neoplasms are composed of parenchymal and stromal regions [38]. The parenchyma consists of transformed neoplastic cells, while the stroma comprises nonneoplastic blood vessels and tissues that provide the parenchymal cells with gas exchange, nutrients and waste disposal [11,39]. The stroma is represented histologically as Vextracellular and physiologically as ve. In our study, histologically determined EES, Vextracellular, obtained by the color-thresholding method was statistically the same as ve measured by DCE MRI. Thus, DCE MRI tumor contrast agent extravasation parameters have a histological correlation. The ve obtained from DCE MRI was selected based on areas of maximal contrast agent transfer (i.e., top five KTRANS/VT), and these values were statistically the same as the values of maximal stromal tissues observed in the histological sections. Thus, these regions may be angiogenically dense, and this may be the physiological reason behind the similarity between permeability-selected ve and Vextracellular. Our data did not show a correlation between KTRANS/VT and Vintracellular. That is, we did not find that high cellularity predicted high contrast agent transfer rates. As KTRANS/VT is a function of both PS and F, perhaps the values we are measuring are not flow limited (i.e., not exclusively permeability dependent and, therefore, not directly related to angiogenesis and its resultant cellular growth). Originally, DCE MRI used whole-lesion averaged KTRANS/VT as a diagnostic tool [32,40–43]. Subsequent

4. Discussion DCE MRI has the potential for differentiating and typing between benign and malignant breast tumors based on their microvasculature characteristics [33,34]. Tumor type is related to neovascularization, and the physiological effects

Table 6 Average and standard deviations of ve, vp and Kp↔t/VT as a function of resolution for noninfiltrating papillary carcinoma (n=3) 469 μm

Average S.D.

938 μm

2500 μm

ve

vp

Kp↔t/VT [L (kg min)−1]

ve

vp

Kp↔t/VT [L (kg min)−1]

ve

vp

Kp↔t/VT [L (kg min)−1]

0.06 0.02

0.002 0.003

0.04 0.03

0.05 0.02

0.004 0.004

0.03 0.02

0.03 0.02

0.001 0.001

0.012 0.008

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

1289

Fig. 13. Diagnostically relevant high-value PS “hot spots” are detected at 469 μm and averaged to lower values at 938 and 2500 μm resolutions. A distribution plot of Kp↔t/VT voxel frequency for 468, 938 and 2500 μm resolution FFT images.

Fig. 14. FFT reconstruction maps of ve (left), vp (center) and Kp↔t/VT (right) at 469 μm resolution.

1290

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293 Table 7 ANOVA of top five PS “hot spots” comparing 469, 938 and 2500 μm resolutions

Fig. 15. FFT reconstruction maps of ve (left), vp (center) and Kp↔t/VT (right) at 938 μm resolution.

studies focused on specific ROI that were most likely more specific for tumor diagnosis based on KTRANS/VT [5,44–47]. The next step was to analyze the smallest ROI possible, namely, individual voxels, and to map the spatially dependent contrast agent characteristics for a tumor. Low-resolution data (e.g., whole tumors, selected ROI and large voxels) have superior quality of fit compared to smaller voxels (higher r2 and lower P values) and would, on first inspection, seem more suited for accurate tumor typing than the high-

Parent resolution (μm)

Compared resolution (μm)

P

469 469 938

938 2500 2500

.0014 b.0001 b.0001

resolution data. However, the low-spatial-resolution data are not as sensitive to tumor microvascular heterogeneity. It does not detect voxel-dependent KTRANS/VT, as well as dynamic high-spatial-resolution images. The top five KTRANS/VT values obtained from 469 and 938 μm resolutions are significantly different from those at the clinical imager resolution (2500 μm; Pb.0001) and from one another (P=.0014). Thus, attempts at using lowerresolution DCE MRI for differential diagnosis may suffer, since they average out the KTRANS/VT values that are the diagnostically important criterion for tumor typing and grading. The top five KTRANS/VT values obtained from MRI spatial resolutions from 469 μm (0.220 mm2) to 938 μm (0.880 mm2) are statistically different from one another and compare with the clinically relevant microscope FOV range from 0.152 mm2 (390 μm diameter) to 0.740 mm2 (860 μm diameter) reported by Weidner et al. [20,21]. Thus, detection of diagnostically relevant KTRANS/VT requires a spatial resolution window of approximately 469–938 μm if accurate differential tumor diagnosis with DCE MRI is to succeed with high specificity.

Fig. 16. FFT reconstruction maps of ve (left), vp (center) and Kp↔t/VT (right) at 2500 μm resolution.

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

Fig. 17. Error maps at 469 μm resolution for P value (3,49) DF (left) and r2 (right) based on FFT reconstruction.

DCE MRI uses physiological information (ve, vp and KTRANS/VT) to infer histological properties and, from that, tumor typing. Most of the inferred histological information in DCE MRI is believed to be within PS [48] and F, both of which are contained within KTRANS/VT; however, as we have demonstrated, some of this information may be also contained within ve and vp. Contrast agent interstitial tumor transit time will alter ve, while angiogenesis may be localized

1291

Fig. 19. Error maps at 2500 μm resolution for P value (3,49) DF (left) and r2 (right) based on FFT reconstruction.

to areas of high ve. Capillary density may be correlated to vp. Thus, to accurately diagnose tumor type and grade, ve, vp and KTRANS/VT must all be considered. The top five KTRANS/VT values and their corresponding ve can statistically differentiate IDCs and noninvasive papillary carcinomas for the 469 and 938 μm resolutions (P=.0017 and P=.0047, respectively), but not for the 2500 μm resolution (P=.9008),

Fig. 18. Error maps at 938 μm resolution for P value (3,49), df (left) and r2 (right) based on FFT reconstruction.

1292

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

Table 8 Comparison of tumor type based top five PS “hot spots” and corresponding ve Resolution (μm)

Tumor type

Tumor type

P

469 938 2500

IDC IDC IDC

Noninvasive papillary carcinoma Noninvasive papillary carcinoma Noninvasive papillary carcinoma

.0017 .0047 .9008

with only five and three tumors, respectively. Based on this spatial-resolution-dependent diagnostic capability, it implies that KTRANS/VT and the associated ve in high-resolution data are the same as the prognostic indicators used in histological tumor specificity determinations, meaning that the highspatial-resolution data are more diagnostically accurate than the low-resolution data. Thus, an appropriate resolution window is essential for DCE MRI tumor diagnosis to achieve extremely high levels of specificity. Specifically for this study within this resolution window, KTRANS/VT values with corresponding ve values, like vascular density “hot spots,” are diagnostic for these two tumor types. However, further studies with greater numbers and a variety of tumors are essential. Acknowledgment This study was supported by PHS grant 1 R01 CA8700901 awarded by the National Institutes of Health, National Cancer Institute. References [1] Su MY, et al. Correlation of dynamic contrast enhanced MRI parameters with microvessel density and VEGF for assessment of angiogenesis in breast cancer. J Magn Reson Imaging 2003;18: 467–77. [2] Esserman L, et al. Contrast-enhanced magnetic resonance imaging to assess tumor histopathology and angiogenesis in breast carcinoma. Breast J 1999;5(1):13–21. [3] Jacobs M, et al. Benign and malignant breast lesions: diagnosis with multiparametric MR imaging. Radiology 2003;229:225–32. [4] Kvistad KA, et al. Breast lesions: evaluation with dynamic contrastenhanced T1-weighted MR imaging and with T*2-weighted first-pass perfusion MR imaging. Radiology 2000;216(2):545–53. [5] Aref M, Brechbiel M, Wiener EC. Identifying tumor vascular permeability heterogeneity with magnetic resonance imaging contrast agents. Invest Radiol 2002;37(4):178–92. [6] Mitchell DG, Cohen MS. MRI principles. Philadelphia: Saunders; 2004. [7] Padhani AR. Dynamic contrast-enhanced MRI in clinical oncology: current status and future directions. J Magn Reson Imaging 2002;16: 407–22. [8] de Jong JS, van Diest PJ, Baak JP. Hot spot microvessel density and the mitotic activity index are strong additional prognostic indicators in invasive breast cancer. Histopathology 2000;36(4):306–12. [9] Weidner N. Tumoral vascularity as a prognostic factor in cancer patients: the evidence continues to grow. J Pathol 1998;184:119–22. [10] Weidner N. Tumour vascularity and proliferation: clear evidence of a close relationship. J Pathol 1999;189(3):297–9. [11] Connolly JL, et al. Principles of cancer pathology. Cancer Medicine. Hamilton (Ontario): BC Decker, Inc.; 2000.

[12] Fox SB. Tumour angiogenesis and prognosis. Histopathology 1997;30: 193–301. [13] Fujimoto K, et al. Small peripheral pulmonary carcinomas evaluated with dynamic MR imaging: correlation with tumor vascularity and prognosis. Radiology 2003;227:786–93. [14] Hanahan D, Weinberg RA. The hallmarks of cancer. Cell 2000;100: 57–70. [15] Matsubayashi R, et al. Breast masses with peripheral rim enhancement on dynamic contrast-enhanced MR images: correlation of MR findings with histologic features and expression of growth factors. Radiology 2000;217:841–8. [16] Nicholson B, et al. Angiogenesis in prostate cancer: biology and therapeutic opportunities. Cancer Metastasis Rev 2001;20:297–319. [17] Eberhard A, et al. Heterogeneity of angiogenesis and blood vessel maturation in human tumors: implications for antiangiogenic tumor therapies. Cancer Research 2000;60:1388–93. [18] Jaeger TM, et al. Tumor angiogenesis correlates with lymph node metastasis in invasive bladder cancer. J Urol 1995;154(1):69–71. [19] Weidner N, Semple JP, Welch WR. Tumor angiogenesis and metastasis — correlation in invasive breast carcinoma. N Engl J Med 1991;324(1):1–8. [20] Weidner N. Intratumor microvessel density as a prognostic factor in cancer. [comment]. Am J Pathol 1995;147(1):9–19. [21] Weidner N, et al. Tumor angiogenesis: a new significant and independent prognostic indicator in early-stage breast carcinoma. J Natl Cancer Inst 1992;1:1875–87. [22] Begum R, et al. Radial intratumoral increase and correlation of microvessels and proliferation in solid breast carcinoma. Histopathology 2002;43:244–53. [23] Stoica G, Koestner A, Capen CC. Neoplasms induced with high single doses of N-ethyl-N-nitrosourea in 30-day-old Sprague–Dawley rats, with special emphasis on mammary neoplasia. Anticancer Res 1984;4 (1–2):5–12. [24] Chandra S, et al. Application of Reduced-encoding Imaging with Generalized-series Reconstruction (RIGR) in dynamic MR imaging. J Magn Reson Imaging 1996;6(5):783–97. [25] Hanson JM, et al. Fast dynamic imaging using two reference images. Magn Reson Med 1996;36:172–5. [26] Jones BA. MATLAB: nlinfit. Natick (Mass): The Mathworks, Inc.; 1994. [27] Byrkit DR. Statistics today: a comprehensive introduction. Menlo Park (Calif): The Benjamin/Cummings Publishing Co., Inc.; 1987. [28] Russo J, et al. Tumours of the mammary gland. In: Tutusov V, Mohr U, editors. Pathology of tumours in laboratory animals, tumours of the rat. New York (NY): Oxford University Press; 1990. [29] Morris TW, Katzberg RW. Intravascular contrast media: properties and effects. In: Katzberg RW, editor. The contrast media manual. Philadelphia: Lippincott, Williams and Wilkins; 1991. [30] Bieser A. A pharmacokinetic analysis of Gd-DTPA in the rat: correlations to a contrast enhancement study. Department of physiology. Urbana (Ill): University of Illinois at Urbana-Champaign; 1995. [31] Brasch RC. Rationale and applications for macromolecular Gd-based contrast agents. Magn Reson Med 1991;22:282–7. [32] Tofts PS, Kermode AG. Measurement of the blood–brain barrier permeability and leakage space using dynamic MR imaging: 1. Fundamental concepts. Magn Reson Med 1991;17:357–67. [33] Daldrup-Link HE, et al. Quantification of breast tumor microvascular permeability with feruglose-enhanced MR imaging: initial phase II multicenter trial. Radiology 2003;229(3). [34] Hayes C, Padhani AR, Learch MO. Assessing changes in tumour vascular function using dynamic contrast-enhanced magnetic resonance imaging. NMR Biomed 2002;15(2):154–63. [35] Baxter LT, Jain RK. Transport of fluid and macromolecules in tumors: III. Role of binding and metabolism. Microvasculature Res 1991;41 (2):252–72. [36] Matsumura Y, Maeda H. A new concept for macromolecular therapeutics in cancer therapy: mechanism of tumoritropic accumulation of proteins

M. Aref et al. / Magnetic Resonance Imaging 26 (2008) 1279–1293

[37] [38]

[39] [40]

[41]

[42]

and the antitumor agent SMANCS. Cancer Res 1986;46(12, Pt 1): 6387–92. Takakura Y, et al. Disposition characteristics of macromolecules in tumor-bearing mice. Pharm Res 1990;7(4):339–46. Uddin R, Saeed IM. Structure and growth of tumors: the effect of Cartesian, cylindrical, and spherical geometries. In: Diller KR, editor. Biotransport: heat and mass transfer in living systems. New York: New York Academy of Sciences; 1998. Russo J, Russo IH. Atlas and histologic classification of tumors of the rat mammary gland. J Mammary Gland Biol Neoplasia 2000;5(2). Tofts PS, Berkowitz B, Schnall MD. Quantitative analysis of dynamic Gd-DTPA enhancement in breast tumors using a permeability model. Magn Reson Med 1995;33(4):564–8. Su M-Y, et al. Characterization of N-ethyl-N-nitrosourea induced benign and malignant mammary tumors in rats by using three contrast agents of various molecular weights. ISMRM, Sydney, Australia; 1998. Su M-Y, et al. Tumor characterization with dynamic contrast-enhanced MRI using MR contrast agents of various molecular weights. Magn Reson Med 1998;39:259–69.

1293

[43] Su M-Y, et al. Characterization of N-ethyl-N-nitrosourea-induced malignant and benign breast tumors in rats by using three MR contrast agents. J Magn Reson Imaging 1999;9:177–86. [44] Brasch R, et al. Assessing tumor angiogenesis using macromolecular MR imaging contrast media. J Magn Reson Imaging 1997;7(1): 68–74. [45] Daldrup H, et al. Correlation of dynamic contrast-enhanced magnetic resonance imaging with histologic tumor grade: comparison of macromolecular and small-molecular contrast media. Pediatr Radiol 1998;28(2):67–78. [46] Daldrup H, et al. Quantification of the extraction fraction for gadopentetate across breast cancer capillaries. Magn Reson Med 1998;40(4):537–43. [47] Liney GP, et al. Dynamic contrast-enhanced MRI in the differentiation of breast tumors: user-defined versus semi-automated region-ofinterest analysis. J Magn Reson Imaging 1999;10(6):945–9. [48] Simpson JF, Dutt P, Page DL. Expression of mitoses per thousand cells and cell density in breast carcinomas: a proposal. Hum Pathol 1992;23 (6):608–11.