While the concept of diffusion MRI emerged in the mid-1980s, diffusion MRI has become a mainstay of modern clinical imaging. Diffusion MRI is both a powerful method and concept because diffusing water molecules provide unique information about the microscopic architecture of tissues. Water diffusion is significantly diminished in most malignant tissues, and diffusion MRI, which requires no tracer injection, is rapidly becoming the modality of choice for detecting, characterizing, or even classifying malignant lesions, especially in the breast (31). Diffusion MRI is deeply rooted in the concept that, during their diffusion-induced movements, molecules probe the structure of tissues at a microscopic scale, well beyond the usual millimeter resolution of images. During typical diffusion imaging times of about 50-100 ms, water molecules move through tissues on average over distances of about 1-15 μm, bouncing off, passing through, or interacting with many tissue components, such as cell membranes, fibers, or macromolecules. Due to the tortuous movement of water molecules around these obstacles (“hindered” diffusion), the actual diffusion distance is reduced compared to free water. Therefore, non-invasive observation of water diffusion-induced displacement distributions in vivo provides unique clues to the fine structural features and geometric organization of cells in tissues, as well as to changes in these features as a function of physiological or pathological states.

MRI signals can be sensitized to diffusion by applying a pair of sharp magnetic field gradient pulses, the duration and separation of which can be adjusted to achieve a specific level of diffusion sensitization defined as the “b-value.” By acquiring data with different gradient pulse amplitudes, images with different degrees of diffusion sensitivity are obtained. The overall effect of diffusion in the presence of these gradient pulses is a signal attenuation and the MRI signal becomes “diffusion weighted”, hence the term “Diffusion Weighted Imaging” (DWI). The signal attenuation is more pronounced when large values of b are used and when diffusion is fast (because molecules diffuse over larger distances) ( Figure 2 ). It is important to note that only the displacement (diffusion) component in the direction of the gradient pulses is detectable, but the diffusion can be anisotropic.

In DWI, qualitative contrast depends not only on diffusion, but also on other MRI parameters, such as T1 and T2 water relaxation times, which can lead to well-known artifacts, such as the “T2-shine-through” effect, as high T2 signal lesions (e.g., necrosis, cysts) can retain a relatively high signal level at high b values. Therefore, these images are often combined numerically to determine a quantitative estimate of the diffusion process in each image location, through an Apparent Diffusion Coefficient (ADC), “apparent” because diffusion is impeded by many processes (33):

where S(b₀) and S(b₁) are the signals (in a voxel or region of interest, ROI) acquired at the b values b₀ and b₁, respectively. This simple ADC is an incredibly robust and powerful parameter, which has been widely used in all clinical applications of diffusion MRI since its inception (34). The optimal value of b₁ that provides the best contrast-to-noise ratio in breast tissue, i.e., sufficient attenuation of the signal by scattering while maintaining a sufficient signal level is about 800s/mm² (13).

Beyond molecular diffusion, blood microcirculation in capillary networks (perfusion) also contributes to the diffusion MRI signal. Indeed, the flow of blood water in pseudo-randomly oriented capillaries (at the voxel level) mimics a random walk (“pseudo-diffusion”) which leads to an attenuation of the signal in the presence of diffusion encoding gradient pulses. This effect has been named IntraVoxel Incoherent Motion (IVIM) (35). In the presence of blood microcirculation, the global attenuation of the MRI signal, S(b)/S(0), becomes the sum of two components, one for tissue diffusion and one for the blood compartment:

where f_IVIM is the fraction of circulating blood, D* is the pseudo-diffusion coefficient attributed to the random microcirculation of blood, D is the diffusion coefficient of water in tissue, and D_blood is the diffusion coefficient of water in blood. The perfusion effect is observed only at low values of b, because the pseudo-diffusion coefficient, D*, associated with blood flow is higher than the water diffusion coefficient and decreases more rapidly with the b-value.

IVIM MRI has become an important modality for perfusion imaging, with applications throughout the body (31, 36), particularly in cancer imaging (detection of neovascularization and treatment efficacy). A key feature of IVIM diffusion MRI is that it does not involve contrast agents, and it may serve as an attractive alternative to perfusion MRI in some patients with contraindications to contrast agents, or in patients with renal insufficiency at risk for NSF (see above).

Another important feature of diffusion MRI, which should be considered, counter-intuitively, as an advantage and not as a limitation, is that the ADC value depends on the acquisition parameters, especially the b-value, because diffusion in tissues is not “free” but “hindered”. With free (or “Gaussian”) diffusion, as in a cyst, the ADC remains the same regardless of the set of b values used to measure it (only the accuracy of ADC estimates changes with b values). However, in most tissues, the ADC value decreases as the diffusion sensitivity is increased by the b value (32) ( Figure 2 ).

The reason is that an increasing number of molecules slowed down by their interaction with microstructural tissue components (fibers, cell membranes) during their diffusion movements become visible in the highly diffusion-sensitized MRI signal. This non-Gaussian diffusion behavior is therefore more pronounced when high b values are used. In short, sticking to the “optimal” b-value (e.g. 800s/mm²) deprives one of the potentially valuable clinical information about tissue microstructure encoded in the “non-Gaussian diffusion” provided by higher b-values. To reveal this hidden information about tissue microstructure, one must rely on models other than the standard ADC. There are essentially two types of such models. Some approaches aim to model the diffusion MRI signal biophysically, based on the different tissue compartments present in the tissue, as with NODDI (Neurite Orientation Dispersion and Density Imaging) used in the brain (37). The other way is simply to model the decay of the scattering signal mathematically, empirically, without any assumptions about the underlying biophysical properties of the tissue. Although several models have been proposed (38), the most popular approach simply quantifies the deviation of the scattering signal behavior from an ideal Gaussian behavior. This is the so-called Kurtosis model (39), also called Diffusion Kurtosis Imaging, DKI (40). With the Kurtosis model, which also includes the IVIM effect, the signal is described as follows:

ADC₀ is the extrapolated ADC value as b approaches 0 and K is the Kurtosis quantifying the deviation from Gaussian scattering (K=0 for Gaussian diffusion). Kurtosis has shown great potential for characterizing pathological or physiological conditions (41). A major drawback of DKI, however, is that it requires the acquisition of large data sets with multiple values of b to be fitted with equation (3), which significantly increases acquisition times, a premium in clinical practice.

However, it is possible to obtain quantitative information about non-Gaussian diffusion with data sets acquired with a limited range of b values. For example, using data acquired for only 2 b values, one can calculate a shifted ADC (sADC). The concept of sADC (31) is based on the use of shifted key b-values (200 and 1500s/mm² for the breast, instead of 0 and 800s/mm²) providing an interesting balance between Gaussian and non-Gaussian diffusion effects. This approach has been evaluated for the breast (42). Another approach, S-index, provides a direct classification of tissue types by calculating a distance between the acquired signals and a library of reference (“signature”) signals from known or simulated tissues (e.g., benign, malignant, etc.) by intrinsically accounting for Gaussian and non-Gaussian diffusion effects, without the need for any mathematical or biophysical modeling (43). This approach has also been shown to provide the immunohistochemical status and molecular subtypes of invasive breast carcinomas (44) ( Figure 3 ).

It is also possible to estimate the main parameters of the IVIM/Kurtosis model described by equation [3], f_IVIM, ADCo and K, without fitting, using a limited set of 4 b values (b0, b1, b2 and b3 in ascending order), providing the signal:noise ratio is not too low. The proposed algorithm assumes that IVIM effects become negligible in signals acquired above b1 and that non-Gaussian diffusion effects appear visible in b2 and b3 signals. According to this 4b algorithm the model parameters estimates can be calculated as:

where D1=ln[S(0)/S(b1)]/(b1-b0); D2=ln[S(0)/S(b2)]/(b2-b0); D3=ln(S(0)/S(b3)/(b3-b0); H=(D2-D3)/(b3-b2)+ln(1-F)/(b3.b2); F= 1-exp[-(D1-D2)(b1.b2)/(b2-b1)]; A= (D2-D3)/(b3-b2)+ln(1- f_IVIM)/(b3.b2).

In the absence of IVIM and non-Gaussian diffusion effects one obviously has D1=D2=D3=sADC, ADCo=sADC and K=0.

If non-Gaussian diffusion is present without IVIM effects (f_IVIM=0) ADCo and K are obtained exactly as:

A graphical interpretation of this set of equations can be given by plotting the (log) of the (curved) signal attenuation versus the b value and the straight lines corresponding to D1, D2, D3 and sADC ( Figure 4 ). In the absence of IVIM and non-Gaussian diffusion effects the signal attenuation follows a straight line with a slope D1=D2=D3=sADC=ADCo. In the presence of IVIM effects only the curvature at low b values creates an angle between the D1 and D2 lines. From this angle f_IVIM can be estimated (Eq. 4) while ADCo remains very close to the sADC (Eq. 6). In the presence of non-Gaussian diffusion only the curvature at high b values forms an angle between the D2 and D3 lines, from which K can be estimated (Eq. 7). However, one can see that f_IVIM and non-Gaussian diffusion slightly contribute also to the angle between the D2 and D3 lines, and the angle between the D1 and D2 lines, respectively. Hence, estimated f_IVIM and K values must be corrected (variables H and A in Eq. 4, 6). Note that the sADC now includes ADCo, f_IVIM and K effects, so that the ADCo values derived from sADC must be corrected (A variable). Also, with this algorithm D* cannot be estimated, however, a review of the literature shows that D* is a parameter difficult to estimate even with the full fitting approach, resulting in extremely variable clinical relevance.

It is expected that those abbreviated quantitative DWI protocols will play a major role, in addition to qualitative DWI, such as DWIBS (see below), in the context of breast cancer screening with diffusion MRI.

Molecular mobility in biological tissues may not be the same in all directions, which is referred to as diffusion anisotropy. In the breast diffusion anisotropy can arise from the geometric organization of the glandular tissue around ducts. To characterize the effects of anisotropy, diffusion-weighted images must be sensitized to diffusion along multiple directions (at least 6) within the Diffusion Tensor Imaging (DTI) (45) With DTI one gets information on the tissue mean diffusivity, MD, which is equivalent to an orientation invariant ADC, and lambda values (λ₁, λ₂, and λ₃) which give diffusivity along the main diffusivity directions (so-called eigenvectors ϵ₁, ϵ₂, and ϵ₃). The eigenvector ϵ₁, associated with the highest λ value, λ₁, is aligned along the main orientation of aligned structures (e.g. ducts), allowing to produce maps showing their orientation in space. Some vendors propose to estimate the MD from a set of 3 orthogonal directions, but this is an approximation that should not be used in the presence of strong anisotropy effects. The genuine mean diffusivity is simply the average of the 3 λ values. The other important parameter, called Fractional Anisotropy (FA), quantifies the degree of anisotropy (FA = 0 indicates that diffusion is isotropic). It is calculated from the λ values. Whereas the existence of diffusion anisotropy in fibroglandular breast tissue has been claimed by many groups (46–49), the nature of the anatomical features which might cause this anisotropy remains somewhat controversial. Some studies have shown that breast cancer lesions could be associated with significantly lower FA values relative to normal breast tissue, and that λ₁ or (λ₁–λ₃) could overperform the ADC (or MD) for lesion detection and classification (50, 51). However, one has to keep in mind that λ₁ and λ₃ (and FA which depends on them) are, by principle, highly sensitive to noise because of the strongly non-linear nature of the DTI calculation algorithm. The mere fact that MD values are lower in malignant lesions than in normal tissue might lead to reduced FA values, which should not necessarily be interpreted as “reduced anisotropy” (32).

Most often diffusion MRI is used qualitatively for lesion detection. Lesion detection can be achieved from DWI acquired with high b values, which have a higher contrast between breast lesions (which appear bright) and normal parenchyma (dark background). Breast cancer detection using DWI has been shown to be more sensitive than MMG, with the DWI screening approach allowing to detect mammographically occult cancers (53–55) and DWI has been shown to detect significantly more contralateral breast cancers in women with unilateral breast cancer than MMG (56). High b-values are also useful in decreasing false-positive breast cancer cases (57).

A variant of the DWI techniques for qualitative lesion detection is DWIBS (Diffusion-weighted Whole-body Imaging with Background body signal). A previous study in 280 patients has shown that the diagnostic performance using non-contrast technique including DWIBS for breast lesion detection (sensitivity, specificity, diagnostic accuracy, Positive Predictive Value (PPV) and Negative Predictive Value (NPV) values of 94%, 79%, 86%, 79% and 94%, respectively) was comparable to that of DCE-MRI (sensitivity, specificity, diagnostic accuracy, PPV and NPV values of 98%, 83%, 90%, 84% and 98%, respectively) (58). DWIBS performed with Maximum Intensity Projection (MIP) mapping also has a comparable diagnostic performance (sensitivity, specificity, PPV and NPV values of 92%, 94%, 93%, and 92%, respectively) to that of DCE-MRI performed with MIP (sensitivity, specificity, PPV and NPV values of 85%, 90%, 89%, and 87%, respectively). MIP-DWIBS has been shown to rule out previously suggested malignancy on screening MMGs in 50 participants with carcinoma in 24 patients (59).

Nevertheless, a unique feature of breast DWI is its quantitative assessment capability. As the most popular quantitative marker, the ADC can be used as a threshold to sort out benign from malignant lesions (13, 60, 61), but also to build a lexicon to describe and classify lesions, for instance to distinguish breast cancers from benign lesions (13). Many groups have also found significant differences of ADC values between benign and metastatic breast lymph nodes (62–64), however, their diagnostic performance in differentiating these lymph nodes still need further investigation (64) compared to simpler markers such as the lymph node size. Quantitative DWI in addition to DCE‐MRI and other plain MRI such as T1WI and T2WI also leads to improved diagnostic performance, in terms of specificity for BI-RADS (Breast Imaging-Reporting And Data System) 3 and 4 lesions, or evaluating malignancies with BI-RADS 4 lesions (65, 66).

DWI is often used in multiparametric protocols in combination with other MRI modalities, such as DCE-MRI, contributing to improve overall diagnostic specificity and accuracy over DCE-MRI alone (67), especially when examining non-Gaussian diffusion (42) ( Figure 5 ). The combination of DCE-MRI and DWI could increase diagnostic accuracy in characterization of non-mass-like enhancement lesions (68). It has also been reported that DWI combined with T2WI improved the diagnostic specificity of enhancing lesions incidentally detected in breast DCE-MRI (69), and that multiparametric DWI outcome parameters have associations with molecular prognostic factors or subtypes (70, 71).

Advanced diffusion markers (IVIM, Kurtosis, DTI) can further increase diagnostic performance, although there are not yet used in routine clinical practice ( Figures 3 , 5 ). IVIM parameters have been shown to provide a high diagnostic performance in differentiating benign and malignant breast tumors (sensitivity = 86%, specificity = 86%, AUC = 0.91 for D, sensitivity = 80%, specificity = 76%, AUC = 0.85 for f, and sensitivity = 84%, specificity = 59%, AUC = 0.71 for D*) (72), especially in combination with DCE-MRI (73), and IVIM parameters are known to be correlated with DCE-MRI parameters (74). IVIM histogram parameters have been shown to be associated with molecular prognostic factors (75, 76). Regarding DKI higher K and lower MD values are usually observed in malignant compared to benign lesions (41, 77), DKI was found to be useful in the differentiation of additional suspicious lesions at preoperative breast MRI (78). In 2 recent meta-analyses of DKI studies (79, 80) the sensitivity and specificity of K and MD to differentiate malignant from brewing breast lesions were found to be around 89-90% and 86-88% for K, and 84-86% and 83-88% for MD. The utility of DKI in differentiating molecular prognostic factors (81) or predicting treatment response (82) has also been reported.

As for DTI parameters, malignant breast lesions have significantly lower MD and λ₁ values compared to benign lesions (46). Indeed, λ₁ and MD are known to have a high diagnostic performance in differentiating malignant and benign breast lesions (AUC 0.97, sensitivity 93%, specificity 92% for λ₁ and AUC 0.92, sensitivity 87%, specificity 83% for D (50), although this trend might not be related to diffusion anisotropy [see above (32)]. Similarly, the use of FA to differentiate between malignant and benign breast tumors remains controversial, suggesting that caution should be exercised in the use of this parameter, although several studies have suggested its usefulness to sort out malignant and benign lesions (14, 50). Yamaguchi et al. (83) also reported higher FA in lesions with more favorable prognostic factors, such as positive estrogen receptor status, lower nuclear grade and cancer intrinsic subtype, and reduced DTI metrics had association with poor prognostic factors of breast cancer (84). A correlation between DTI parameters and molecular prognostic factors (estrogen receptor status or Ki-67) has been shown (49) and DTI has been investigated to differentiate recurrent breast cancer from post operative changes with breast-conserving surgery in patients (85).

In reviewing the literature, the overall sensitivity of DWI alone is very high, approaching 90% for detection of breast malignancy (with a specificity of approximately 82%, superior to any other imaging modality, including GBCA MRI) (51).

Many investigators have studied the potential of DWI alone for non-contrast cancer detection. The performance of DWI imaging for cancer detection is variable across studies, with a mean sensitivity of 81% (range 44-97%) and a mean specificity of 88% (range 73-96%) (86). This variation could be due to the diversity of the study population as well as the image acquisition protocols, highlighting the need for standardization (see below). Nonetheless, DWI based primarily on qualitative assessment is less sensitive than DCE MRI (mean sensitivity of 80 vs. 90s % for DWI vs. DCE MRI in studies (54, 58, 59, 67, 87–93). This situation is entertaining the idea that diffusion MRI would be difficult as a stand-alone modality compared to DCE-MRI and multiparametric MRI (67). However, in reviewing the literature, the overall sensitivity of quantitative ADC alone is very high, approaching 90% for detection of breast malignancy (with a specificity of approximately 82%, superior to any other imaging modality, including GBCA-MRI) (51). In any case, with respect to screening, diffusion MRI offers much better clinical performance than MMG or ultrasound (94). This is an important point, considering that repeated use of GBCA would be a problem for screening. Not only does DWI remain more sensitive than MMG across studies (52, 92, 95, 96), but mammography-occluded breast cancers are better represented with DWI than with ultrasound (94).

Single-shot echo-planar imaging (EPI) is currently the method of choice for in vivo diffusion imaging, as it allows efficient and ultrafast acquisition of multiple diffusion-weighted images (different b-values) without in-plane motion artifacts, to which diffusion MRI is notoriously sensitive. Nevertheless, EPI has several limitations related to spatial resolution, artifacts, and signal-to-noise ratio. In particular, small breast lesions (<2 mm) may be undetectable. In addition, EPI requires a very homogeneous magnetic field. For breast imaging, field inhomogeneities may be more pronounced at the air/tissue interface in the anterior part of the breast, resulting in local image distortion or signal loss. Another source of geometric distortion comes from eddy currents induced by the switching of strong diffusion encoding gradient pulses. Therefore, the degree of geometric distortion increases with the b-value. This geometric distortion must be corrected before performing any quantitative analysis involving multiple values of b to avoid artifacts around small lesions, especially at high spatial resolution. Segmented EPI acquisitions (e.g., ‘RESOLVE’ (Readout Segmentation of Long Variable Echo-trains) (105) can overcome these limitations at the cost of longer diffusion times and a sensitivity to motion between acquired segments that must be corrected using ad-hoc approaches during image reconstruction. Parallel acquisition techniques, which allow simultaneous signal collection using an array of multiple RF coils, can also address these limitations. Incorrect fat suppression can also lead to misinterpretation of diffusion MRI, as residual fat present in breast tumors results in low diffusion values, mimicking malignancy, visually and quantitatively (ADC values). The Spectrally Adiabatic Inversion Recovery (SPAIR) method has been recommended for breast imaging (13).

Efforts are also needed on the image processing side. Diffusion-weighted images are often noisy, especially for high b-values, because the signal is strongly attenuated by the diffusion effect. Noise is a vicious enemy because it is not always visible, while having a profound impact on the values of the parameters estimated with the various models available including ADC. For high b-values, due to the nature of the MRI signal (a “magnitude” signal that cannot be negative), there always remains a background noise signal and the diffusion signal remains above a threshold, the “noise floor”, instead of asymptotically approaching 0, resulting in underestimated ADC values. If one classifies lesions (e.g., benign or malignant) on the basis of ADC threshold values, it is easy to see that this trap of underestimated ADC could lead to a significant bias toward the “malignant” nature of lesions. Therefore, an adequate signal-to-noise ratio must be ensured, e.g., by increasing the voxel size (at the expense of spatial resolution) or by repeating image acquisitions at high b-values for signal averaging before amplitude reconstruction (which unfortunately increases acquisition time). Finally, background noise effects must also be removed from the signals before analysis, especially in images acquired at high b-values (106). Noise effects may partly explain the discrepancies in the literature on the different reported values of diffusion MRI and IVIM parameters. Image preprocessing could also include steps to correct for motion artifacts and geometric distortion before the signals can be processed to calculate ADC values or estimate parameters for advanced DWI models. Another problem with clinical diffusion MRI is that quantitative analysis is often performed remotely on workstations and not on the acquisition console, which is cumbersome. Efforts are underway by vendors to provide dedicated tools for breast DWI [see the final chapters of the book (10)].

DWI data analysis would also benefit from recent developments in artificial intelligence (AI). Various approaches are being investigated for breast MRI, as well as remarkably increasing applications of convolutional neural network models (107) and machine learning (108). For example, a recent study showed that DWI radiomic classifiers for differentiating suspicious lesions in 50 asymptomatic women screened with MMG outperformed the average ADC, with an area under the curve (AUC) of receiver operating characteristics (ROC) of 84.2%/85.1% for unconstrained/constrained radiomic classifiers compared with 77.4% for the average ADC (109). The AI-based multiparametric MRI approach, including DCE, T2WI, and DWI, had better diagnostic performance (AUC ROC area of 0.852) than ultrafast DCE alone (0.811) (110). Machine learning with multiparametric MRI (DCE, DWI, and T2WI) also found that several features, including those of DWI (minimum ADC), were relevant features for predicting residual cancer burden (111). Whole breast segmentation on DWI data from different institutions and scanner types was also found to be effective using deep learning methods, which could facilitate computer-assisted quantitative analyses of DWI images of the breast (112).