Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
ORIGINAL ARTICLE Quantitative Non-Gaussian Diffusion and Intravoxel Incoherent Motion Magnetic Resonance Imaging Differentiation of Malignant and Benign Breast Lesions Mami Iima, MD, PhD,* Kojiro Yano, MD, PhD, MA,*† Masako Kataoka, MD, PhD,* Masaki Umehana, BSc,‡ Katsutoshi Murata, BSc,§ Shotaro Kanao, MD,* Kaori Togashi, MD, PhD,* and Denis Le Bihan, MD, PhDk ¶ Objectives: The purpose of this study was to explore the potential of nonGaussian diffusion and perfusion magnetic resonance imaging (MRI) using intravoxel incoherent motion (IVIM) MRI for the diagnosis of breast lesions. Materials and Methods: This study included 26 women with breast lesions. Diffusion-weighted images were acquired using 16 b values up to 2500 s/mm2 and analyzed using a kurtosis diffusion model (apparent diffusion coefficient [ADC0] and kurtosis [K]) for the diffusion component and IVIM model (perfusion fraction [fIVIM] and pseudodiffusion coefficient [D*]) for the perfusion component. Diagnostic performance of diffusion and perfusion parameters was evaluated from receiver operating characteristic analyses. Results: The ADC0 in malignant lesions was significantly lower than that in benign lesions and normal tissue (P < 0.001, P < 0.001), whereas K was significantly higher (P < 0.05, P < 0.001), as well as fIVIM (P < 0.05, P < 0.01). No significant difference was found in D*. The receiver operating characteristic analysis gave high area under the curve values for ADC0, K, and fIVIM for distinguishing malignant from benign lesions (0.99, 0.85, and 0.82, respectively). The ADC0 allowed benign tumors to be identified with 100% negative predictive value and malignant tumors with 100% sensitivity. The malignant/benign diagnosis thresholds were 1.4 10−3 mm2/s as well as 0.6 and 7%, respectively, for ADC0, K, and fIVIM. Conclusions: With a proper methodological framework, IVIM MRI can provide valuable information on tissue structure and microvasculature beneficial for the diagnosis of breast cancer lesions. Key Words: diffusion-weighted MR imaging, intravoxel incoherent motion, kurtosis, breast cancer, differential diagnosis, noise correction (Invest Radiol 2015;50: 205–211) he concept behind diffusion magnetic resonance imaging (MRI)1 is to probe tissue microstructure through the measurement of water molecular displacements powered by Brownian motion. As such, water diffusion is becoming an important biomarker of cancer because it has been observed that the water apparent diffusion coefficient (ADC) is significantly reduced in many primary or secondary cancer tissues.2 A putative mechanism for this drop in ADC is the increase in the T Received for publication May 5, 2014; and accepted for publication, after revision, July 31, 2014. From the *Department of Diagnostic Imaging and Nuclear Medicine, Kyoto University Graduate School of Medicine, Kyoto; †Information Science and Technology, Osaka Institute of Technology, Osaka; ‡Kyoto University Faculty of Medicine, Kyoto; §Siemens Japan K.K., Tokyo; kHuman Brain Research Center, Kyoto University Graduate School of Medicine, Kyoto, Japan; and ¶NeuroSpin, CEASaclay, Gif-sur-Yvette, France. Supported by a Grant-in-Aid for Scientific Research from the Ministry of Education, Culture, Sports, Science and Technology of Japan (No. 24591756); a Grand-inAid for Japan Society for the Promotion of Science Fellows from Japan Society for the Promotion of Science (No. 25-2430); and the Louis D. Foundation of the Institute of France. Conflicts of interest and sources of funding: none declared. Supplemental digital content is available for this article. Direct URL citation appears in the printed text and is provided in the HTML and PDF versions of this article on the journal’s Web site (www.investigativeradiology.com) Reprints: Denis Le Bihan, MD, PhD, Bât 145, Point Courrier 156 F-91191 Gif/Yvette, France. E-mail: [email protected]. Copyright © 2014 Wolters Kluwer Health, Inc. All rights reserved. ISSN: 0020-9996/15/5004–0205 Investigative Radiology • Volume 50, Number 4, April 2015 density of cell membranes (caused by cell proliferation), which act as obstacles for water diffusion.3 Such interactions of diffusing water with tissue components are responsible for a non-Gaussian diffusion behavior (as opposed to free diffusion), which has been readily observable when using a high degree of diffusion weighting (the so-called high b values) now achievable on clinical MRI scanners.4,5 Another aspect of diffusion MRI is its sensitivity to perfusion because the flow of blood water in randomly oriented capillaries mimics a diffusion process through the “intravoxel incoherent motion” (IVIM) effect.6 Recently, IVIM MRI has undergone a striking revival, especially for body organ studies.4 Intravoxel incoherent motion MRI might be especially useful to investigate cancer tissues for which vascularity is a key parameter, to characterize tumors and predict or monitor therapeutic responses. A key feature of IVIM diffusion MRI is that it does not involve contrast agents and can therefore be an alternative for perfusion MRI in patients exposed to the risk for nephrogenic systemic fibrosis. Indeed, a significant progress has been made since the early introduction of the IVIM and ADC concepts4,5 and important methodological issues must now be revisited to exploit the full potential of quantitative diffusion and IVIM MRI in clinical practice. Hence, our aim was to explore the potential of non-Gaussian diffusion and perfusion MRI using IVIM MRI using an updated quantitative imaging framework to simultaneously take into account IVIM effects (low b values) and non-Gaussian diffusion effects (high b values) as well as issues related to effects of noise. This framework is introduced in the context of breast cancer. The choice of breast cancer, one of the most common cancers in women worldwide, was motivated by the high potential of diffusion MRI to differentiate benign and malignant breast lesions, to evaluate tumor extension,7–9 and to predict the response to neoadjuvant chemotherapy in patients with breast cancer.10 MATERIALS AND METHODS Patient Population This retrospective study was approved by an institutional review board, and informed consent was waived. Between April and December of 2011, a total of 26 consecutive patients were enrolled in this study, with the following inclusion criteria: breast lesions with a long diameter larger than 8 mm at the time of MRI examination. Four patients were excluded because of poor signal quality caused by either motion artifacts or noise. Thus, 22 patients (31–74 years; mean, 52.4 years) were consequently selected for the study, with 23 lesions: 15 malignant (9 invasive ductal carcinomas, 4 invasive lobular carcinomas, 1 mucinous carcinoma, and 1 phyllodes tumor) and 8 benign tumors (2 fibroadenomas, 4 fibrocystic changes, 1 pseudoangiomatous stromal hyperplasia, and 1 inflammation). One patient had bilateral lesions: an invasive ductal carcinoma in 1 breast and a fibrocystic lesion in the other breast. Standard of Reference Malignant tumors were histopathologically diagnosed through biopsy first and confirmed after surgery. Benign lesions were diagnosed through biopsy first and the absence of tumor growth during their follow-up period for at least 18 months by ultrasonographic or www.investigativeradiology.com Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. 205 Investigative Radiology • Volume 50, Number 4, April 2015 Iima et al radiological findings.11 Breasts without abnormal findings were regarded as negative for cancer on the basis of clinical and/or radiological follow-up for at least 8 months. All pathologic results were defined according to the World Health Organization classification of breast tumors.12 MRI Acquisitions Breast MRI was performed at 3 T (Trio B17; Siemens Medical Solutions) using a dedicated 16-channel breast array coil (6 channels covering each breast13). The following images were obtained after localizers were acquired: (1) bilateral fat-suppressed T2-weighted images (repetition time/echo time, 5500/77 milliseconds; flip angle, 140 degrees; field of view, 330 330 mm2; matrix, 448 448; slice thickness, 1 mm; acquisition time, 1 minute and 30 second); (2) trace-weighted diffusion images (single-shot echo planar imaging) with spectral attenuated inversion recovery for fat suppression with the following parameters: b values (3, 5, 10, 20, 30, 50, 70, 100, 200, 400, 600, 800, 1000, 1500, 2000, 2500 s/mm2) (the minimum b value was 3 s/mm2 owing to the presence of crusher pulses); repetition time/echo time, 4,600/86 milliseconds; flip angle, 90 degrees; field of view, 160 300 mm2; matrix, 80 166; slice thickness, 3.0 mm; 25 slices without gap; bandwidth, 1585 Hz; acquisition time, 3 minutes 55 seconds; generalized autocalibrating partially parallel acquisitions with acceleration factor of 2; (3) T1-weighted images were obtained using a 3-dimensional fat-suppressed gradient-echo sequence (repetition time/echo time, 3.7/1.36 milliseconds; flip angle, 15 degrees; field of view, 330 330 mm2; matrix, 346 384; slice thickness, 1 mm; acquisition time, 60 seconds). Fat-suppressed T1-weighted dynamic contrast-enhanced images were also acquired before and after injection (0–1, 1–2, and 5–6 minutes) of a gadolinium-based contrast agent (ProHance; Eisai, Tokyo, Japan) but were not considered for this study. Data Processing Traditionally, magnetic resonance images have been quantitatively processed using fitting algorithms that provide estimate of parameters according to a given nonlinear signal model.14 Intravoxel incoherent motion/diffusion MRI has been no exception. Although the original IVIM model6 remains somewhat unchallenged, several models have been successfully introduced to take into account the non-Gaussian (not free) water diffusion behavior observed at high b values in biological tissues (see Supplemental Digital Content 1, http://links.lww.com/RLI/A169). Here, we have chosen the kurtosis model because it seems more robust when using medium-range b values (lower than 3000 s/mm2).15 The overall measured signal, M(b), can then be modeled as follows (see Supplemental Digital Content 1, http://links.lww.com/RLI/A169): h i1=2 MðbÞ ¼ S ðbÞ2 þNCF (½1) n h io with S ðbÞ ¼ So f IVIM expðbDÞþ ð1f IVIM Þ exp bADC 0 þ ðbADC 0 Þ2 K=6 (½2) where S0 is the theoretical signal acquired at b = 0; fIVIM, the (T1-,T2weighted) volume fraction of incoherently flowing blood in the tissue; D*, the pseudodiffusion coefficient associated to the IVIM effect; and ADC0, the virtual ADC that would be obtained when b approaches 0. The dimensionless coefficient K (kurtosis) characterizes the degree of deviation of the signal behavior from a monoexponential decay (K = 0 when the diffusion-driven molecular displacements obey the Gaussian law), a marker of the heterogeneity of the diffusion environment. The noise correction factor (NCF) is a parameter that characterizes the “intrinsic” noise contribution from the image acquisition setup (depending on the coils, the MRI sequence parameters, etc) (see Supplemental Digital Content 1, http://links.lww.com/RLI/A169). 206 www.investigativeradiology.com To overcome some pitfalls often encountered with the usual data-fitting algorithms (see Discussion), we have introduced a completely new approach. Instead of fitting the signal data to the IVIM/ diffusion model using the standard iterative (fitting) search approach, we directly derive the parameters by comparing the raw signal data with those of a database of simulated signals built once for all for the entire study using an exhaustive set of discretized and bounded parameter combinations with Eq. [2]. In other words, a distance, di, is calculated between the measured signal attenuation profile, M(b), and each simulated signal, Sdb(i), of the database. The parameter combination, Pi, giving the shortest distance, dmin, is deemed to represent the searched parameter estimates. In our study, the distance, di , was defined as follows: di ¼ S b ðMðbÞ Sn2dbðiÞ ðbÞ S 2 ðb ¼ 0Þ þ NCF 1=2 ÞwðbÞ ½ 2 (½3) where Sndb(i)(b) is the (normalized) simulated signal from the database for the parameter combination P i(S i0, f iIVIM, D*i, ADC i0, Ki) obtained from Eq. [2], scaled with S(b = 0), which was estimated as [M2(b = 3) – NCF]1/2. W(b) is a weighting factor introduced to compensate the acquisition sampling bias, at the disadvantage of the diffusion parameters, when the number of data points with high b values, nbhigh, is lower than those with low b values, nblow, (as in our study). This bias can be seen as a trend (slope) in the error residuals plotted against b values between the simulated and raw signals (good fitting should lead to a random distribution of the error residuals). W(b) was set to 1/nblow (1/9) for signals acquired with b < 400 mm2/s and 1/nbhigh (1/7) for signals acquired with b ≥ 400 mm2/s. Above b = 400 mm2/s, the IVIM contribution to the signal becomes negligible16 (as a worst-case scenario, maximizing IVIM contribution with fIVIM = 20% and D* = 8.0 10−3 mm2/s, the IVIM contribution at b = 400 s/mm2 is 0.8%, well within noise). The noise correction factor has been added to the simulated signals and not subtracted from the measured signals (inverting Eq. [1]) to circumvent potential problems that could arise with negative squared values when signals are close to the noise floor (high b values). The database was built with the following parameter ranges: S0: [0.975–1.025] step 0.002 fIVIM: [0–20] step 1 (%) D*: [3.5–25] step 0.5 (10−3 mm2/s) ADC0: [0.8–2.1] step 0.1 (10−3 mm2/s) K: [0.0–1.3] step 0.1 To limit its size, we excluded from the database unrealistic parameter combinations leading to Sdb(i)(b = 2500) > Sdb(i)(b = 2000) (which may arise when both ADC0 and K are high, a known pitfall of the kurtosis model), or D* < 3ADC0 (lower limit to allow a meaningful separation of 2 exponentials in Eq. [2]6). Furthermore, the steps within each parameter range were set to provide a reasonable accuracy on the estimated parameters while restricting the size of the data bank. The method was implemented with MATLAB code (Mathworks, Naticks, MA) to run the analysis at the region of interest (ROI) level and at the pixel level so as to generate parametric maps of the diffusion and perfusion parameters. To show pitfalls that may result from using a standard (monoexponential) ADC diffusion model and compare our results with those in the literature, we also performed a 2-step process at the ROI level using the nonlinear subspace trust region fitting algorithm built into MATLAB (Mathworks, Natick, MA). The ADC and fIVIM values (referred thereafter as ADCmono and fIVIMmono to distinguish them from the ADCo and fIVIM values obtained with the full approach described previously) were also estimated by fitting the diffusion signal decay, S/S0, for 200 s/mm2 ≤ b ≤ 1000 s/mm2 with a monoexponential function, then by fitting the residual of the signal at b < 200 s/mm2 with © 2014 Wolters Kluwer Health, Inc. All rights reserved. Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. Investigative Radiology • Volume 50, Number 4, April 2015 Quantitative Diffusion/IVIM MRI in Breast Cancer TABLE 1. Diffusion Parameters in Malignant lesions, Benign Lesions, and Normal Tissue P* Parameter ADC0 (10−3 mm2/s) K ADCmono (10−3 mm2/s) Malignant Benign Normal 1.05 (0.94–1.17) 0.82 (0.70–0.94) 1.00 (0.88–1.11) 1.73 (1.51–1.94) 0.55 (0.42–0.68) 1.50 (1.35–1.66) 1.97 (1.87–2.07) 0.25 (0.16–0.34) 1.78 (1.67–1.90) Malignant Versus Benign Tumor Malignant Versus Normal Tissue Benign Versus Normal Tissue <0.001 <0.05 <0.001 <0.001 <0.001 <0.001 <0.05 <0.01 <0.05 Values are presented as mean (95% confidence interval). *Bonferroni corrected. ADC0 indicates apparent diffusion coefficient; K, kurtosis. the IVIM model (equations A1, A2, A3; supplemental information, Supplementary Digital Content 1, http://links.lww.com/RLI/A169), as also done by other groups17–19: Regions of Interest Two readers (M.I., radiologist A, and M.K., radiologist B, with 6 years and 10 years of experience in breast MRI, respectively, blinded to the final pathologic results) manually drew ROIs on the slice with the largest tumor area using the b = 0 and b = 1000 s/mm2 diffusionweighted images, avoiding T2-shine through areas usually found in necrotic or cystic parts under guidance of the T2-weighted images. The ROIs were defined as slightly smaller than the actual lesions to reduce partial volume effects. The ROIs were also drawn in the normal homogeneous breast parenchyma for all patients as controls (avoiding contamination by fatty tissue) except in 2 patients who had bilateral lesions or surgical history in the controlateral breast. The median and range of ROI size were 97.6 (63.4–196.0) mm2, 75.9 (42.2–263.3) mm2, and 171.7 (119.3–271.6) mm2 for the malignant tumors, benign tumors, and normal tissue, respectively. Statistical Analysis All the parameters in malignant and benign lesions as well as normal tissue were compared with the Mann-Whitney test. Bonferroni correction was used to account for multiple comparisons. Receiver operating characteristic curve analyses were conducted to assess ADC0, K, and fIVIM in terms of their utility for discrimination of malignant and benign lesions. An optimal threshold was established for those parameters, giving the best sensitivity and specificity balance. For all tests, a P < 0.05 was considered statistically significant. All statistical analyses were conducted using statistical software MedCalc (version 11.3.2.0, Mariakerke, Belgium). RESULTS The diffusion parameters across malignant and benign tumors as well as normal breast tissue are summarized in Table 1. The ADC0 in malignant lesions was significantly lower than that in benign lesions (P < 0.001) and normal breast tissue (P < 0.001). The K in malignant lesions was significantly higher than that in benign lesions (P < 0.05) and normal breast tissue (P < 0.001). Benign tumors had significantly lower mean values of ADC0 and higher mean values of K than normal breast tissue did (P < 0.05 and P < 0.01, respectively). Although ADCmono values obtained with a monoexponential model also showed significant difference between malignant and benign tumors or normal tissue (P < 0.001, P < 0.001) or that between benign tumors and normal tissue (P < 0.05), they were always found smaller than ADC0 values, reflecting improper handling of non-Gaussian diffusion. In malignant tumors, fIVIM was significantly higher than that in benign tumors and normal breast (P < 0.05 and P < 0.01, respectively) (Table 2). The fIVIMmono values obtained using the 2-step process ADC model were underestimated in malignant lesions and overestimated in benign lesions and normal tissue, underlining the fact that the bias in fIVIM estimates using a simple ADC model for diffusion depends on the degree of non-Gaussian diffusion effects. There was no significant difference in fIVIMmono or D* across malignant and benign tumors and normal breast tissue. Examples of parametric diffusion and perfusion maps obtained in the patients with invasive ductal carcinoma and pseudoangiomatous stromal hyperplasia (PASH) are shown in Figures 1 and 2. The pattern differences observable between the fIVIM, ADC0, and K maps are striking, showing tumor heterogeneity not detectable at the ROI level. Locations presenting a high fIVIM/low ADC0/high K combination are potentially the most active parts, suggesting spots where biopsy should be made. The receiver operating characteristic analysis gave high AUC values for AUC of ADC0, K and fIVIM for distinguishing malignant TABLE 2. IVIM Parameters in Malignant and Benign Lesions and Normal Tissue P* Parameter fIVIM, % D* (+) (10−3 mm2/s) fIVIMmono, % Malignant Benign Normal Malignant Versus Benign Tumor Malignant Versus Normal Tissue Benign Versus Normal Tissue 12.3 (8.86–15.7) 10.9 (5.95–15.9) 8.63 (7.39–9.86) 5.00 (0.60–9.40) 13.9 (5.96–21.9) 8.66 (6.97–10.4) 4.45 (2.63–6.27) 12.6 (8.78–16.3) 6.64 (4.97–8.31) <0.05 0.52 1.00 <0.01 0.36 0.20 1.00 1.00 0.44 Values are presented as mean (95% confidence interval). *Bonferroni corrected. fIVIM indicates perfusion fraction; D*, pseudodiffusion coefficient. © 2014 Wolters Kluwer Health, Inc. All rights reserved. www.investigativeradiology.com Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. 207 Investigative Radiology • Volume 50, Number 4, April 2015 Iima et al FIGURE 1. Invasive ductal carcinoma in a 61-year-old woman. A, Anatomical contrast-enhanced image. B, The fIVIM map. C, The ADC0 map. D, The K map. The white rectangle in A shows the area covered by the parametric maps. The high fIVIM fraction area at the periphery of the tumor in B matches the enhancing lesion very well in A. The lesion center has low perfusion (suggesting necrosis). An area on the left part of the tumor exhibits a low ADC0 with high K value, suggesting high cellularity (viable malignant component), whereas the central part has a high ADC0 and low K, suggesting lower cellularity (possible necrosis). from benign lesions. Although those AUC values were not tested for statistical significance between them given the small size of the patient population, the AUC of ADC0 (0.99) was found higher than that of K and fIVIM (0.85 and 0.82, respectively) (Table 3). The ADC0 allowed benign tumors to be identified with 100% negative predictive value and malignant tumors with 100% sensitivity. The malignant/benign diagnosis thresholds were 1.4 10−3 mm2/s as well as 0.6 and 7%, respectively, for ADC0, K, and fIVIM. DISCUSSION The diffusion and perfusion parameters obtained from the kurtosis diffusion/IVIM model could well differentiate malignant from benign breast tumors with high sensitivity and specificity. The ADC0 values in malignant lesions were significantly lower than those in benign lesions and normal tissues, consistent with the ADC values reported from other studies,17,19–21 but our ADC0 values (obtained with a larger range of b values and using a non-Gaussian diffusion model) were lower than ADCmono values and ADC values reported in the literature. Indeed, with state-of-the-art MRI scanners, it is now becoming possible to extract further useful information from IVIM MRI, acquiring data beyond the usual range of b values used in clinical practice (600–1000 s/mm2).16,19 The ADC values, as obtained from only 2 b values or a monoexponential diffusion model, are significantly lower than the ADC0 value estimated when higher b values are taken into account using non-Gaussian diffusion model such as the kurtosis model, depending on tissue types, which points out that some important information may be missed when using a simple ADC.5 Indeed, significant differences for K between malignant, benign, and normal breast tissue were found, as also observed in prostate cancer.22,23 The failure to take into account non-Gaussian diffusion effects furthermore results in ADC values that are highly dependent on the choice of b values,19,21 making it difficult to compare studies across centers. In contrast, an important feature of ADC0 and K is that their intrinsic values do not depend on the b values used for image acquisition (only the accuracy on their estimates will still depend on the number and ranges of b values). Other non-Gaussian models, such as the biexponential model, were also explored in this study, according to Iima et al.18 However, the results were not as robust compared with the kurtosis model (not shown) when dealing with noisy clinical data. Indeed, because there are only 2 unknown parameters to estimate, the kurtosis model is gaining momentum in clinical studies.22–24 Whereas ADC0 represents diffusion at low b values, K comes mainly from the curvature of the diffusion signal decay observed at high b values. The ADC0 is considered to reflect more diffusion in the extracellular space, which also reflects the amount of cell filling (shape and size) in tissues and cell proliferation. Large K values point out to enhanced diffusion hindrance effects in malignant tissues, likely related to cell proliferation and membrane interactions with diffusing water. Another important point is that using a simple ADC (calculated from 2 b values or using a monoexponential fit) to remove diffusion effects from the overall signal to extract IVIM parameters, as often performed,16,17,19 may not be sufficient. The curvature from non-Gaussian diffusion not taken into account with a simple ADC model propagates at low b values and results in a pseudo-IVIM effect, leading to an overestimation of fIVIM and D* values (see supplemental information, Supplemental Digital FIGURE 2. Pseudoangiomatous stromal hyperplasia in a 46-year-old woman. A, Anatomical contrast-enhanced image. B, The fIVIM map. C, The ADC0 map. D, The K map. The white rectangle in A shows the area covered by the parametric maps. The left-center lesion with a high ADC0 (C) and a very low K (D) corresponds to the nonenhancing lesion in A. The corresponding lesion in B has heterogeneous fIVIM fraction. Such parameter combination highly suggests the presence of a tissue structure containing parts with a quasi-free diffusion component (fluid-filled lobules or cystic component, although a detailed correlation between pathologic and radiologic images was not performed for this retrospective study) coexisting with more cell-filled parts with higher and more heterogeneous fIVIM values. 208 www.investigativeradiology.com © 2014 Wolters Kluwer Health, Inc. All rights reserved. Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. Investigative Radiology • Volume 50, Number 4, April 2015 Quantitative Diffusion/IVIM MRI in Breast Cancer TABLE 3. Diagnostic Abilities of Diffusion and Perfusion MRI Parameters Parameter AUC Threshold Sensitivity, % Specificity, % PPV, % NPV, % Accuracy, % ADC0 (10−3 mm2/s) K fIVIM 0.99 0.85 0.82 ≦1.40 >0.60 >7.0 100 (15/15) 80 (12/15) 80 (12/15) 88 (7/8) 88 (7/8) 88 (7/8) 94 (15/16) 92 (12/13) 92 (12/13) 100 (7/7) 70 (7/10) 70 (7/10) 96 (22/23) 83 (19/23) 83 (19/23) ADC0 indicates apparent diffusion coefficient; AUC, area under the curve; fIVIM, perfusion fraction; K, kurtosis; MRI, magnetic resonance imaging; NPV, negative predictive value; PPV, positive predictive value. Content 1, http://links.lww.com/RLI/A169 and Fig. 3, especially when large bvalues are used to calculate the ADC, as also reported by Bokacheva et al.19 This bias might partially explain why the D* values reported in this study are somewhat lower than the values reported by other groups.16,19 Note that caution should be advised because poor fat suppression may also result in high K values owing to the fact that fat tissues have very low diffusion coefficients.25 Furthermore, noise bias effects at high b values, resulting from the non-Gaussian nature of the noise in magnitude-reconstructed images26 must be addressed to avoid erroneous K value estimates. The main effect of such noise is that it may mimic a curvature in the diffusion signal attenuation plot because, at high b values, the signal reaches a “noise floor” and does not get to 0 (see Supplemental Digital Content 1, http://links.lww.com/RLI/A169, Figs. 4 and 5. The signal attenuation appears curved, even for monoexponential diffusion, and fitting signals with diffusion models will give erroneous values (eg, underestimation of ADC0, overestimation of K). Many groups have investigated the effect of non-Gaussian noise in diffusion MRI and suggested methods to retrieve signal values from noise-corrupted FIGURE 3. Simulated data without tissue perfusion. Even in the absence of perfusion, owing to the curvature of the diffusion signal attenuation (crosses), S/So, fitting the diffusion signal decay with a straight line (ADC model, here from signal values at b = 200 and 1000 s/mm²) automatically places data points at low b values above the ADC line (dashed line), resulting in a false IVIM effect (non-0 value for fIVIM in the absence of perfusion). © 2014 Wolters Kluwer Health, Inc. All rights reserved. data.27–31 In this study, we have used a simple procedure where a noise correction factor is experimentally obtained through a phantom calibration process relying on the diffusion MRI signal property itself. Another issue is that the usual fitting approaches (such as the popular Marquardt-Levenberg algorithm) used to estimate diffusion and perfusion parameters from measured signals suffer from several drawbacks. Such algorithms are generally very sensitive to noise.14 This sensitivity leads to instabilities in the estimated parameter values, especially when many parameters are set free, as with Eq. [2]. To mitigate this issue, the IVIM/diffusion equation is often fitted in 2 steps: first, estimating the diffusion parameters (usually from a monoexponential diffusion model), then estimating the perfusion parameters from the FIGURE 4. Noise correction modeling in 3 alkanes. The raw signal attenuation (crosses) is well fitted by the diffusion/noise model (Eq. [A7]) (dashed line). After noise removal, the corrected signal points (circles) are perfectly aligned, as expected for free, Gaussian diffusion. The differences between raw and corrected signal values are large for nonane, which has a low signal at large b value (high diffusion), but very small for undecane, which keeps a high signal level at high b values (low diffusion). Error bars are smaller than the dot size and not shown. www.investigativeradiology.com Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. 209 Investigative Radiology • Volume 50, Number 4, April 2015 Iima et al FIGURE 5. A, Left, Without noise correction, the raw signal (crosses) appears well fitted with the kurtosis diffusion model (circles), but the IVIM component (residual signal after removal of the diffusion component) would be negative (inset). B, Right, The raw signal (crosses) is, however, very much better fitted by the noise correction model (dashed line). The corrected signal attenuation (circles) becomes close to a straight line, which is confirmed as K becomes close to 0, whereas ADC0 value decreases to 2.07 × 10−3 mm²/s. The IVIM effect (inset) is now positive, as expected (fIVIM = 6.8%), and well fitted by an exponential decay (D* = 14.7 × 10−3 mm²/s). residual signal.17–19 This 2-step process may increase robustness (because there are fewer parameters to estimate for each step) and is based on the assumption that IVIM perfusion effects are not expected to contribute to the signal for b values above a cutoff value. However, it is preferable to handle Eq. [2] as a whole because diffusion effects are obviously present at low b values. Another drawback is that local minima may result in parameter estimates that are somewhat far from the true values, so that the results are very sensitive to the set of initial parameter values that are required to launch the fitting process. The exhaustive search approach that we have introduced not only alleviates the issues of local minima and sensitivity to initial values of the iterative approach but is also much more efficient (hence, faster) in terms of computing requirements because only a simple distance needs to be calculated, whereas the iterative method requires a bunch of more complex calculus elements (such as those present in Eq. [2]) to be performed for each iteration (or even twice if a 2-step approach is used to sequentially estimate diffusion and perfusion parameters). Interestingly, because the noise correction factor may vary in space across the image (see Supplemental Digital Content 1, http://links.lww.com/RLI/A169), the noise correction factor may also be included as 1 of the free parameters to estimate within the database. In the future, several improvements can be made to the approach. We have observed (data not shown) that some parameter combinations, although associated with the shortest distance with the measured signal, could sometimes not properly reflect the signal model, as given by Eq. [2], resulting in a trend (slope) for the error residual plotted against b values. The reason for this bias is that biexponential fitting is sensitive to the sampling acquisition scheme: if the number of data points at low and high b values is not balanced (which could be justified, given that IVIM effects are small) “blind” fitting may favor the IVIM component of the signal. In this study, we have weighted the distance by the number 210 www.investigativeradiology.com of low/high b value signals to successfully overcome this problem. Other distance definitions to minimize this bias will have to be investigated in the future, for instance, weighing the distance calculated for each b value in Eq. [3] by the signal amplitude or by the interval between b values to homogenize acquisition sampling biases. One of our limitations is the small number of the population size; the IVIM and diffusion parameter thresholds for the best diagnostic performance will likely change a little bit when using a larger cohort of patients. Furthermore, the patients were referred for an MRI examination because suspicious lesions had been found with mammography or ultrasound examinations. Hence, there is a prevalence bias toward patients with high risk for breast cancer, which could impact our statistical results for predictive values (underestimation of NPV and overestimation of PPV). This pitfall is common to most MRI studies of breast cancer because MRI is far too expensive to be used as a screening modality at this stage. Possible effects of anisotropy such as those in the canals32 were also not investigated because only diffusion-trace–weighted images were acquired. Motion correction and registration of images across b values have been found beneficial for brain studies33 but are precluded for breast studies because of the absence of fixed, visible anatomical landmarks, which are necessary for image realignment, especially on the diffusionweighted images where contrast varies deeply across b values. A point regarding the interpretation of the fIVIM parameter is that it remains T1- and T2-weighted. Conversion of fIVIM to a flowing blood volume34 would require the removal of relaxation effects, which is not straightforward.35 Relaxation parameters depend on the field strength, the tissue type, and, for the blood, the arterial/vein ratio.36–38 Further studies comparing IVIM data with results of quantitative dynamic contrast-enhanced perfusion MRI models39 or arterial spin labeling methods40 may be of great interest, as well as refining the IVIM © 2014 Wolters Kluwer Health, Inc. All rights reserved. Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. Investigative Radiology • Volume 50, Number 4, April 2015 exponential model used in Eq. [2], depending on the underlying vascular functional architecture (see Appendix).6 No significant difference could be found in D* between tissue type, as observed in some previous IVIM study.17 Finally, morework is needed to precisely model the effects of blood microcirculation on the IVIM signal and to establish a relationship between IVIM parameters and tissue perfusion, such as blood volume and flow.34 In conclusion, we have shown that multiple diffusion and perfusion MRI parameters can be obtained in a clinical setting with IVIM MRI, provided that an adequate methodological framework is used to correct for noise effects to take into account the non-Gaussian diffusion signal decay. Setting aside the usual data-fitting process with its known sensibility to noise and initial parameter values due to the presence of local minima also appears as an important methodological shift, not only within the scope of diffusion and perfusion MRI. With the exhaustive search method, we have introduced that diffusion and perfusion parameters can be estimated at once (in 1 step) without worrying about those limitations and in a much less computer-intensive manner. Processing speed may also be increased by limiting the size of the database, choosing parameter value ranges according to organ and tissue types. In the context of breast cancer, the diffusion and IVIM parameters, ADC0, K, and fIVIM, may help improve diagnostic accuracy and guide biopsy location. Although these preliminary results need to be validated at a broader scale, they suggest that images of tissue structure and blood microvasculature can be obtained without contrast agents using IVIM MRI, which is an interesting alternative to perfusion MRI. With further study, these applications might play an important role in the screening, diagnosis, or monitoring of the breast cancer as well as serve as a potential prognostic biomarker of breast. ACKNOWLEDGMENTS The authors thank Masakazu Toi, MD, PhD, for his excellent advice to this study; Takuma Imakita, BSc, for his exceptional contribution to software development; as well as Masayuki Nakagawa, RT, and Tosiaki Miyati, PhD, for their excellent contribution in data acquisition. REFERENCES 1. Le Bihan D, Breton E, Lallemand D, et al. MR imaging of intravoxel incoherent motions: application to diffusion and perfusion in neurologic disorders. Radiology. 1986;161:401–407. 2. Türkbey B, Aras Ö, Karabulut N, et al. Diffusion-weighted MRI for detecting and monitoring cancer: a review of current applications in body imaging. Diagn Interv Radiol. 2012;18:46–59. 3. Le Bihan D. The “wet mind”: water and functional neuroimaging. Phys Med Biol. 2007;52:R57–R90. 4. Le Bihan D. Intravoxel incoherent motion perfusion MR imaging: a wake-up call. Radiology. 2008;249:748–752. 5. Le Bihan D. Apparent diffusion coefficient and beyond: what diffusion MR imaging can tell us about tissue structure. Radiology. 2013;268:318–322. 6. Le Bihan D, Breton E, Lallemand D, et al. Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging. Radiology. 1988;168:497–505. 7. Woodhams R, Matsunaga K, Iwabuchi K, et al. Diffusion-weighted imaging of malignant breast tumors: the usefulness of apparent diffusion coefficient (ADC) value and ADC map for the detection of malignant breast tumors and evaluation of cancer extension. J Comput Assist Tomogr. 2005;29:644–649. 8. Partridge S, DeMartini W, Kurland B, et al. Differential diagnosis of mammographically and clinically occult breast lesions on diffusion-weighted MRI. J Magn Reson Imaging. 2010;31:562–570. 9. Iima M, Le Bihan D, Okumura R, et al. Apparent diffusion coefficient as an MR imaging biomarker of low-risk ductal carcinoma in situ: a pilot study. Radiology. 2011;260:364–372. 10. Park SH, Moon WK, Cho N, et al. Diffusion-weighted MR imaging: pretreatment prediction of response to neoadjuvant chemotherapy in patients with breast cancer. Radiology. 2010;257:56–63. 11. Eby PR, DeMartini WB, Gutierrez RL, et al. Characteristics of probably benign breast MRI lesions. Am J Roentgenol. 2009;193:861–867. 12. Tavassoli FA, Devilee P, eds. World Health Organization Classification of Tumours: Pathology and Genetics of Tumours of the Breast and Female Genital Organs. Lyon, France: IARC Press; 2003. © 2014 Wolters Kluwer Health, Inc. All rights reserved. Quantitative Diffusion/IVIM MRI in Breast Cancer 13. Wichmann T, Kurth R, Geppert C, et al. A 16 channel phased array coil optimized for diagnostic breast imaging. In: Proceedings of the 17th Annual Meeting of ISMRM; 2009; Honolulu, HI. Abstract 2969. 14. Gill P, Murray W. Algorithms for the solution of the nonlinear least-squares problem. SIAM J Numer Anal. 1978;15:977–992. 15. Yablonskiy DA, Sukstanskii AL. Theoretical models of the diffusion weighted MR signal. NMR Biomed. 2010;23:661–681. 16. Liu C, Liang C, Liu Z, et al. Intravoxel incoherent motion (IVIM) in evaluation of breast lesions: comparison with conventional DWI. Eur J Radiol. 2013;82:e782–e789. 17. Sigmund E, Cho G, Kim S, et al. Intravoxel incoherent motion imaging of tumor microenvironment in locally advanced breast cancer. Magn Reson Med. 2011;65:1437–1447. 18. Iima M, Reynaud O, Tsurugizawa T, et al. Characterization of glioma microcirculation and tissue features using intravoxel incoherent motion magnetic resonance imaging in a rat brain model. Invest Radiol. 2014;49:485–490. 19. Bokacheva L, Kaplan JB, Giri DD, et al. Intravoxel incoherent motion diffusionweighted MRI at 3.0 T differentiates malignant breast lesions from benign lesions and breast parenchyma [published online ahead of print November 22, 2013]. J Magn Reson Imaging. 20. Woodhams R, Matsunaga K, Kan S, et al. ADC mapping of benign and malignant breast tumors. Magn Reson Med Sci. 2005;4:35–42. 21. Bogner W, Gruber S, Pinker K, et al. Diffusion-weighted MR for differentiation of breast lesions at 3.0 T: how does selection of diffusion protocols affect diagnosis? Radiology. 2009;253:341–351. 22. Rosenkrantz AB, Sigmund EE, Johnson G, et al. Prostate cancer: feasibility and preliminary experience of a diffusional kurtosis model for detection and assessment of aggressiveness of peripheral zone cancer. Radiology. 2012;264:126–135. 23. Pang Y, Turkbey B, Bernardo M, et al. Intravoxel incoherent motion MR imaging for prostate cancer: an evaluation of perfusion fraction and diffusion coefficient derived from different b-value combinations. Magn Reson Med. 2013;69:553–562. 24. Lu H, Jensen JH, Ramani A, et al. Three-dimensional characterization of nonGaussian water diffusion in humans using diffusion kurtosis imaging. NMR Biomed. 2006;19:236–247. 25. Baron P, Dorrius MD, Kappert P, et al. Diffusion-weighted imaging of normal fibroglandular breast tissue: influence of microperfusion and fat suppression technique on the apparent diffusion coefficient. NMR Biomed. 2010;23:399–405. 26. Henkelman RM. Measurement of signal intensities in the presence of noise in MR images. Med Phys. 1985;12:232–233. 27. Gudbjartsson H, Patz S. The Rician distribution of noisy MRI data. Magn Reson Med. 1995;34:910–914. 28. Brion V, Poupon C, Riff O, et al. Parallel MRI Noise Correction: an Extension of the LMMSE to Non Central χ Distributions. Medical Image Computing and ComputerAssisted Intervention—MICCAI 2011. Berlin, Heidelberg: Springer; 2011:226–233. 29. Constantinides CD, Atalar E, McVeigh ER. Signal-to-noise measurements in magnitude images from NMR phased arrays. Magn Reson Med. 1997;38:852–857. 30. Aja-Fernández S, Tristán-Vega A, Hoge WS. Statistical noise analysis in GRAPPA using a parametrized noncentral Chi approximation model. Magn Reson Med. 2011;65:1195–1206. 31. Koay CG, Basser PJ. Analytically exact correction scheme for signal extraction from noisy magnitude MR signals. J Magn Reson. 2006;179:317–322. 32. Eyal E, Shapiro-Feinberg M, Furman-Haran E, et al. Parametric diffusion tensor imaging of the breast. Invest Radiol. 2012;47:284–291. 33. Jenkinson M, Bannister P, Brady M, et al. Improved optimization for the robust and accurate linear registration and motion correction of brain images. Neuroimage. 2002;17:825–841. 34. Le Bihan D, Turner R. The capillary network: a link between IVIM and classical perfusion. Magn Reson Med. 1992;27:171–178. 35. Lemke A, Laun FB, Simon D, et al. An in vivo verification of the intravoxel incoherent motion effect in diffusion-weighted imaging of the abdomen. Magn Reson Med. 2010;64:1580–1585. 36. Lu H, Clingman C, Golay X, et al. Determining the longitudinal relaxation time (T1) of blood at 3.0 Tesla. Magn Reson Med. 2004;52:679–682. 37. Uludağ K, Müller-Bierl B, Uğurbil K. An integrative model for neuronal activityinduced signal changes for gradient and spin echo functional imaging. Neuroimage. 2009;48:150–165. 38. Brown R, McGorty K, Moy L, et al. Sub-millimeter breast imaging and relaxivity characterization at 7T. In: Proceedings of the 19th Annual Meeting of ISMRM; 2011; Montreal, Canada. Abstract 3092. 39. Makkat S, Luypaert R, Sourbron S, et al. Quantification of perfusion and permeability in breast tumors with a deconvolution-based analysis of second-bolus T1DCE data. J Magn Reson Imaging. 2007;25:1159–1167. 40. Buchbender S, Obenauer S, Mohrmann S, et al. Arterial spin labelling perfusion MRI of breast cancer using FAIR TrueFISP: initial results. Clin Radiol. 2013;68:123–127. www.investigativeradiology.com Copyright © 2015 Wolters Kluwer Health, Inc. All rights reserved. 211