Perfusion-driven Intravoxel Incoherent Motion (IVIM) MRI in Oncology: Applications, Challenges, and Future Trends

Abstract

Recent developments in MR hardware and software have allowed a surge of interest in intravoxel incoherent motion (IVIM) MRI in oncology. Beyond diffusion-weighted imaging (and the standard apparent diffusion coefficient mapping most commonly used clinically), IVIM provides information on tissue microcirculation without the need for contrast agents. In oncology, perfusion-driven IVIM MRI has already shown its potential for the differential diagnosis of malignant and benign tumors, as well as for detecting prognostic biomarkers and treatment monitoring. Current developments in IVIM data processing, and its use as a method of scanning patients who cannot receive contrast agents, are expected to increase further utilization. This paper reviews the current applications, challenges, and future trends of perfusion-driven IVIM in oncology.

Introduction

Intravoxel incoherent motion (IVIM) was defined in 1986 by Le Bihan et al.¹ as the “translational movements which within a given voxel and during the measurement time present a distribution of speeds in orientation and/or amplitude”. Such movements correspond mainly to molecular diffusion, but also to microcirculation of blood in pseudo-randomly oriented capillary vessels and even to tissue vibrations used for MR elastography.² The apparent diffusion coefficient (ADC) obtained with diffusion MRI was already shown to include a perfusion related component, especially when very low b-values are used.¹ However, it was shown in 1988³ that perfusion-related effects and diffusion-related effects could be separated within the IVIM composite signal, providing ad hoc acquisition and processing parameters were used. Diffusion MRI is implemented on most clinical MRI systems and today widely used in clinics, and the utility of the ADC in oncology has been very well documented for various organs.^4,5 The potential of perfusion-related IVIM MRI has; however, taken more time to be appreciated, but has enjoyed a significant revival over the past 10 years, as shown in Fig. 1. The reason for perfusion-driven IVIM MRI to lag in time is that IVIM effects are small, requiring very good image quality. With the recent developments in MR hardware and software, which have improved signal-to-noise ratio in IVIM MRI,⁶ it is now possible to obtain reliable estimation of perfusion related IVIM parameters. An important feature of IVIM MRI is that it can provide quantitative information on microcirculation without the use of contrast agents, an important advantage in terms of cost, acquisition times and applicability to patients who cannot receive gadolinium-based contrast agents for different reasons. IVIM MRI can provide quantitative maps of the density of small, functional blood vessels (related to flowing blood volume fraction) (Fig. 2), a key component of angiogenesis.⁷

Fig. 1.

Number of IVIM publications in 1986–2019.

Fig. 2.

Changes in f_IVIM with tumor growth. f_IVIM increases with the proliferation of neovascularity in some tumors. As the pseudo-diffusion coefficient associated with blood microcirculation is much larger than the true diffusion coefficient in tissues, fast-decaying signal (IVIM effect) appears at low b-values. f_IVIM, the flowing blood fraction, increases with the proliferation of neovascularity in some tumors.

Indeed, the microvasculature of tumors, which often exhibits multiple structural and functional abnormalities, is a major target of oncology treatments. Thus, perfusion imaging has become an important means for the management of cancer, whether for diagnosis, characterization, or staging of malignant tumors associated with active angiogenesis. It is also useful in assessing the response to treatment and detection of recurrence.⁸

Thus far, contrast-enhanced (CE) MRI using gadolinium-based contrast media has been a used as the standard imaging method to assess perfusion in different organs (i.e., brain, spine, abdomen, breast, and heart),⁹ because of its better performance compared with other imaging techniques such as CT or ultrasound. CE MRI is highly accurate in evaluating the response to treatment, or assessing for residual disease after neoadjuvant chemotherapy (NAC) of breast cancer.^10–14 The utility of gadoxetic acid as a liver-specific contrast medium is recognized for the diagnosis of hepatic lesions.¹⁵ In addition to assessment of vascularity, more advanced CE MRI methods [e.g. dynamic susceptibility contrast (DSC) and dynamic contrast-enhanced (DCE) imaging] provide quantitative and functional information on the tumor microenvironment (e.g. vascular permeability and blood flow).⁹ However, the emergence of contraindications such as nephrogenic systemic fibrosis (NSF) in patients with impaired renal function, as well as reports of gadolinium deposition in the brain and other tissues^16–18 have raised issues, especially when repeated exams are necessary, as in monitoring treatment response and active surveillance for recurrence in oncology.

In this article, after providing a short summary of IVIM MRI principles, the literature showing what perfusion-related IVIM MRI provides in addition to diffusion MRI (and the ADC) for oncologic applications is reviewed.

IVIM Principles

IVIM MRI is sensitive to both molecular diffusion in tissues and to microcirculation (perfusion) based on the assumption that the flow of blood through capillaries mimics a diffusion process, due to the pseudo-random organization of capillaries in tissue. Microcirculation contributes greatly to the diffusion-weighted MRI signal, S(b), together with genuine water molecule diffusion in tissues:

$$\frac{S(b)}{S_0}=f_{\text{IVIM}}\text{exp}\left(-b\left(D_{\text{blood}}+D^{*}\right)\right)+\left(1-f_{\text{IVIM}}\right)\mathit{Fd}$$ (1)

where f_IVIM is the flowing blood fraction, D^* is the pseudo-diffusion coefficient associated with blood microcirculation, D_blood is the water diffusion coefficient in blood, and Fd, the diffusion-related signal attenuation. Diffusion and perfusion effects can be disentangled from the overall diffusion/IVIM MRI signal because the pseudo-diffusion coefficient associated with blood microcirculation is about ten times larger than the water diffusion coefficient in tissues.

In a simple model assuming diffusion is quasi-Gaussian in tissues (which is only valid at low b-values, in general <600 s/mm² in the body) one has Fd = exp(−bD) where D is the water diffusion coefficient in tissues. In this case, Eq. (1) has a biexponential shape with a “fast” pseudo-diffusion coefficient (D^* + D_blood) and a “slow” diffusion coefficient (D). f_IVIM is sometimes called “f” or “fp,” D^* is sometimes referred to as “Dp,” “ADC_fast,” or “ADC_high,” while D is sometimes called “Dt,” “ADC_slow,” or “ADC_low,” not to be confused with the standard ADC, whose calculation includes both perfusion-related and genuine diffusion effects.¹

However, diffusion signal behavior is non-Gaussian in tissues (especially in cancers with high diffusion hindrance or restriction due to cell proliferation), and fitting IVIM signals into Eq. (1) when including moderate or high b-values results in artificially high f-values. In those conditions, a different function, Fd, must be used to take into account non-Gaussian diffusion effects, as the IVIM/diffusion signal attenuation curve is no longer biexponential (for more details, please see “Challenges”).

Perfusion MRI Methods in Oncology

Angiogenesis plays an important role in the growth of tumors.¹⁹ Tumor vasculature largely consists of immature and tortuous vessels.²⁰ The IVIM framework may be able to capture data on microperfusion in many tumors, as has been validated in several studies (please see next sections).

The main MRI perfusion techniques to date are DSC, DCE, and arterial spin labeling (ASL). Both DSC and DCE require administration of gadolinium, while ASL is contrast-free. DSC is widely used for the clinical evaluation of stroke, tumors, and myocardial ischemia, which involves an intravenous bolus of gadolinium chelate and the serial measurement of signal loss by the passage of the bolus on T₂- or ${\text{T}}_2^{*}$-weighted images.²¹ ASL is mainly used to evaluate blood flow in the brain, heart, kidney, and muscle, and uses the magnetically labeled blood water itself as an endogenous tracer.²² DCE is widely used for the diagnosis or evaluation of treatment response of tumors, where T₁-weighted images (T₁WI) are acquired dynamically before, during, and after bolus injection of a contrast agent (CA) over approximately 5–10 min to allow visualization of arterial and venous phases, including the portal venous phase in the liver. Gadolinium contrast accumulates within the extracellular space during this time frame, and the signal intensity measurements extract quantitative parameters that reflect tissue perfusion, extravascular extracellular space, and vessel permeability.²³

Correlation of IVIM parameter f with tumor blood volume obtained with DCE-MRI has been reported in cancers of the head and neck,²⁴ cervix,²⁵ and some soft tissue cancers;²⁶ no correlation has been found in glioma.²⁷ IVIM reflects all randomly flowing blood in each voxel, while DCE mostly measures CA extravasation, where most of the recorded signals derive from the contrast material accumulating in the interstitial space, so their values will differ.

Validation with Histology

The vasculature considered in IVIM imaging has incoherent flow. Thus, one might expect correlation between IVIM parameters (f or f·D^*) and microvasculature histology, such as microvessel area or microvessel density, especially in tumors. There have been several studies published both in humans and animals, as shown in Table 1.^28–38 It is interesting to see that the correlation can be observed both in human tumors and animal xenografts; however, some studies have found no significant correlation. Bakke et al.³² found no significant correlation between f and microvessel density or vessel size in 12 rectal cancer patients, and Li et al.³⁸ also found no significant correlation between f and microvessel density in 16 liver tumors in rabbits (r = 0.281, P = 0.291). Conversely, IVIM parameter maps obtained by clustering approaches with Gaussian mixture models might be useful for the identification of tumor subregions with proliferative activity (Ki-67 index).³⁹ Still, the correlation between IVIM parameters (f or f·D^*) and microvascular histology is not entirely clear, and further studies are needed to validate the correlation, both in humans and animals.

Table 1.

f, D* or f·D* validation studies with histologic correlation

Humans	Year	Cancer	Subjects (n)	Correlation	Correlation coefficient
Bäuerle et al.28	2013	Rectal cancer without therapy	12	f vs. microvessel area	r = 0.60, P < 0.05
Bäuerle et al.28	2013	Rectal cancer after chemoradiotherapy	9	f vs. microvessel area	r = −0.44, P = 0.29
Klau et al.29	2015	Pancreatic adenocarcinoma and PNET	36 and 6	f vs. microvessel density	r = 0.85, P < 0.01
Surov et al.30	2017	Rectal cancer	17	f vs. microvessel area	r = 0.68, P = 0.003
Togao et al.31	2018	Meningioma	29	f vs. microvessel density	r = 0.69, P < 0.0001
Bakke et al.32	2019	Rectal cancer	12	f·D* vs. microvessel density or vessel size	No significant correlation
Kikuchi et al.33	2019	Pediatric intracranial tumors	17	f vs. microvessel density	r = 0.832, P < 0.0001

Animals	Year	Cancer	Subjects (n)	Correlation	Correlation coefficient
Iima et al.34	2014	Glioma in rats	14	f vs. microvessel density	r = 0.56, P < 0.05
Lee et al.35	2014	Colorectal cancer in mice	25	f, D* vs. microvessel density	f: r = 0.75, P < 0.001
					D*: r = 0.78, P < 0.001
Joo et al.36	2014	Liver tumors in rabbits	21	f, f·D* vs. microvessel density	f: r = 0.52, P = 0.02
					fD*: r = 0.62, P = 0.003
Yang et al.37	2017	Hepatocellular carcinoma mouse model not treated	15	f vs. microvessel density	r = 0.57, P = 0.009
Yang et al.37	2017	Hepatocellular carcinoma mouse model treated	15	f vs. microvessel density	r = 0.44, P = 0.054
Li et al.38	2018	Liver tumors in rabbits	16	f, D* vs. microvessel density	No significant correlation

Several papers have reported the correlation between D and cellularity, however, the correlation between ADC and cellularity has been already extensively investigated in various tumors.

Clinical Applications of IVIM in Oncology

IVIM for characterization and prognostication of tumors

Breast

The diagnostic performance of IVIM parameters for distinguishing between malignant and benign breast tumors has been reported in eight studies.^40,41 Malignant lesions showed significantly lower D in all eight investigations, and seven of the studies demonstrated significantly higher f_IVIM values in malignant lesions.⁴⁰

Lower ADC values in estrogen receptor (ER) or progesterone receptor (PgR) positive tumors have already been reported in many papers.^42–61 Kawashima et al.⁵³ demonstrated significantly lower D and ADC values in luminal B compared with luminal A tumors. Iima et al.⁵⁸ reported significantly lower sADC_200–1500 values in PgR expression. IVIM parameters’ correlation with hormone receptors is also worthy of investigation. IVIM histogram analysis revealed a significant correlation of fp and the pseudo-diffusion coefficient (Dp) with hormone receptor expression (ER or PgR),⁴⁹ while Kim et al.⁵⁰ found that Dp negatively correlated with ER and PgR expression.

Some researchers have also investigated the association of IVIM parameters with pathological biomarkers. Lee et al.⁵¹ found that D_slow 50th, 75th, and 90th percentile values were decreased in ER-positive tumors, and f skewness increased in Ki-67 positive tumors. Suo et al.⁵² reported that the ER expression significantly correlated with ADC, D and f, and D^* significantly correlated with Ki-67 expression. Significantly lower D^* values in borderline and malignant phyllodes tumors compared with fibroadenomas have been observed, which might reflect a slow blood velocity and have some association with quantity of stroma.⁶²

Recent publications have shown the utility of IVIM for monitoring treatment response. Several researchers have reported the utility of D or f for detecting pCR (pathological complete response) after neoadjuvant treatment of breast cancers.^63–65 Che et al.⁶³ demonstrated pretreatment f-value of pCR group significantly higher than that of non-pCR, and Bedair et al.⁶⁴ reported pretreatment diffusion coefficients of pCR group significantly lower than that of non-pCR. Cho et al.⁶⁵ showed that histogram metrics of pseudodiffusion Dp significantly differed between response evaluation criteria in solid tumors responders from nonresponders, while ADC or Dt parameters did not.

Brain tumors

Federau et al.⁶⁶ showed that f-values positively correspond with glioma grade. However, a recent meta-analysis of nine studies in grading gliomas showed higher D^* and lower D in high-grade compared to low-grade gliomas, but no correlation of f with grade.⁶⁷ Puig et al.⁶⁸ have reported correlation of f and D^* values with cerebral blood flow in glioblastomas, and f > 9.86% and D^* > 21.712 × 10⁻³ mm²/s were the thresholds for lower 6-month survival, with both 100% sensitivity and area under the curve (AUC) of 0.893 and 0.857, respectively. Federau et al.⁶⁹ have further reported on f and ADC values for predicting survival in gliomas, suggesting higher f (>0.112), lower ADC (<1033 × 10⁻⁶ mm²/s) and higher relative cerebral blood volume (>3.01) as the indicators of poorer prognosis, with AUCs predicting a 2-year survival of 0.84 for f-value, 0.86 for ADC value, and 0.76 for relative cerebral blood volume.

Head and neck

A recent meta-analysis of the diagnostic performance of combined IVIM parameters in distinguishing among squamous cell carcinomas, lymphomas, malignant salivary gland tumors, Warthin tumors, and pleomorphic adenomas yielded a sensitivity of 85–87% and specificity of 80–100%.⁷⁰ One example of IVIM parameters in characterizing perfusion and diffusion properties of head and neck tumors is shown in Fig. 3.⁷¹ Several studies have found significantly smaller D- and f-values in lymphomas compared with squamous cell carcinomas.^72–75 D in malignant salivary gland tumors was also found to be significantly lower than in pleomorphic adenomas, and significantly higher than in Warthin tumors.^72–74 Different numbers and combinations of b-values have been used in the head and neck, with a median of 10.5 b-values, from 0–800 to 0–1000 s/mm².⁷⁰ Significantly higher f and lower D-values in primary tumors compared with metastatic nodes have been shown in head and neck cancer.⁷⁶ Liang et al. demonstrated that D·D^* is the most significant predictor of lymph node metastasis in head and neck squamous carcinoma.⁷⁷ Fujima et al.⁷⁸ reported that D- and K-values estimated using a hybrid IVIM and diffusion kurtosis imaging (DKI) model were also found useful in predicting future distant metastasis in head and neck squamous cell carcinoma patients.

Fig. 3.

Representative contrast-enhanced T₁-weighted MR images (CE T₁WI) with (a, e, m, q and u) or without (i) fat-suppression, and PP, D, and TIC maps of a squamous cell carcinoma (SCC) node (a–d), lymphoma (e–h), carcinoma ex pleomorphic adenoma (i–l), pleomorphic adenoma (m–p), Warthin tumor (q–t), and schwannoma (u–x). (a–d) A 75-year-old man with an SCC node at level IIA (a), PP = 0.057 (b), D = 0.830 × 10⁻³ mm²/s (c), TIC profile = Type 2 (d). (e–h) A 79-year-old man with lymphoma in the oropharynx (e), PP = 0.04 (f), D = 0.435 × 10⁻³ mm²/s (g), TIC profile = Type 3. (i–l) A 59-year-old man with carcinoma ex pleomorphic adenoma in the palate (i), PP = 0.217 (j), D = 1.005 × 10⁻³ mm²/s (k), TIC profile = Type 3 (l). (m–p) A 34-year-old woman with pleomorphic adenoma in the left parotid gland (m), PP = 0.099 (n), D = 1.287 × 10⁻³ mm²/s (o), TIC profile = Type 2 (p). (q–t) A 63-year-old woman with Warthin tumor in the left parotid gland (q), PP = 0.227 (r), D = 0.485 × 10⁻³ mm²/s (s), TIC profile = Type 4 (t). (u–x) A 24-year-old man with a schwannoma (u), PP = 0.294 (v), D = 1.550 × 10⁻³ mm²/s (w), TIC profile = Type 2 (x). White demarcations on the PP and D maps indicate tumor areas. Adapted from Sumi et al.⁷¹

Monitoring IVIM MRI parameters during treatment (pattern of low pre-treatment D or f-values and an increase in D during treatment) was found to be useful in predicting response to NAC in head and neck cancers, with 64–94% sensitivity and 72–89% specificity.⁷⁰

Liver

Intravoxel incoherent motion has a potential role in staging liver fibrosis, with a sensitivity of 71–81% and specificity of 77–84%, and D has been reported to be significantly lower in malignant compared with benign hepatic tumors.⁷⁹ Although there is evidence of an association of D-values with the histological grade of hepatocellular carcinoma (HCC), attempts at finding an association between f and D^* measurements have been inconclusive, and their added value is still controversial.⁸⁰ Scholars have investigated the variability of IVIM in grading HCC lesions, depending on the fitting methods or ROI positioning.^81,82 Ichikawa et al. demonstrated that the choice of fitting methods affects IVIM parameter values, with smaller D-values found in poorly differentiated, compared with well-to-moderately differentiated HCC, using all methods.⁸¹ Wei et al. reported that the effect of different ROI positioning approaches on IVIM and ADC values is significant in HCC lesions, with ADC_slow most predictive of grading HCC (a negative correlation with HCC grade) (Fig. 4).⁸² Recent investigations have reported the superiority of D or its histogram analysis over ADC in evaluating microvascular invasion in HCC.^83,84

Fig. 4.

MR images in a 58-year-old woman with surgically proven HCC of E–S grade 3. (a) Portal venous phase image. (b) apparent diffusion coefficient (ADC) map. (c) ADC_slow map. (d) ADC_fast map. (e) and (f) map. (f) Intravoxel incoherent motion (IVIM) and DWI fitting of the diffusion signal decay. The tumor demonstrates a slightly high signal intensity on T₂-weighted images and blue areas were observed on the ADC and ADC_slow map, which indicate a poorly differentiated hepatocellular carcinoma. The IVIM-DWI model achieved significantly better fitting than the DWI. Adapted from Wei et al.⁸²

D^* was reported to be more accurate than ADC in distinguishing responders from non-responders to loco-regional treatment in HCCs;⁸⁵ and f-values significantly increased at 2-week follow-up of HCCs responding to sorafenib, while no significant difference was found in ADC or D.⁸⁶ However, another group later found no significant differences in f between responders and non-responders.⁸⁷

Pancreas

Among IVIM parameters, several studies reported that f-value was decreased in pancreatic ductal adenocarcinoma (PDAC) compared with normal pancreatic tissue,^88–93 while no significant difference in D-value was found between carcinoma and healthy tissue,^92,93 except some publications.^89,90 f was found to be most useful for the distinction between pancreatic cancer and chronic pancreatitis, with a trend toward higher f-value in chronic pancreatitis than pancreatic cancer.^88,92,94 D- and f-values have also been found useful to distinguish well/moderately differentiated PDAC from poorly differentiated PDAC, with lower D-values and higher f-values in well/moderately differentiated compared with poorly differentiated PDAC.⁹⁵ Several groups have reported that pancreatic neuroendocrine tumors show higher f-values than PDAC, which is considered to reflect their hypervascularity.^29,88,89,92

Prostate

Several studies have examined the utility of prostate IVIM DWI (D or ADC) in distinguishing prostate cancer from benign hyperplasia and normal tissue, revealing conflicting results in the f measurements in malignant and normal tissue.^96,97 The characteristics of IVIM DWI and MR perfusion parameters in prostate tumors and normal tissues have also been investigated.^97,98

Pang et al.⁹⁷ reported that f is significantly increased (7.2% vs. 3.7%) in tumors compared with normal tissues, in accordance with the volume transfer constant (K_trans; 0.39 vs. 0.18/min) and plasma fractional volume (vp; 8.4% vs. 3.4%). Recently, Beyhan et al.⁹⁸ also found that the mean values of perfusion parameters obtained from the Tofts model [blood and tissue (K_trans), contrast agent back-flux rate constant (K_ep), extravascular extracellular fractional volume (V_e), initial area under curve (iAUC) and χ²] and f significantly increased, and the mean values of Dp and Dt significantly decreased in malignant lesions compared with normal tissue.

Intravoxel incoherent motion histogram metrics might be also useful in the pathological grading of prostate cancers, and it was found that D outperformed conventional ADCs in discriminating low-grade from high-grade prostate cancers.⁹⁹

Female pelvis

The IVIM model has also been intensively investigated in cervical cancer. Poorly differentiated cervical cancer was found to have a lower D than well/moderately differentiated cervical cancer.^100,101 Several studies have reported lower f-values in cervical squamous cell carcinoma (SCC),^101–103 and the results are mixed regarding value for grading of the cervical tumors.^101,103 Interestingly, the f-values at the periphery of cervical cancers were useful in distinguishing tumor grade with higher f in higher-grade tumors,¹⁰⁰ and IVIM histogram metrics distinguished between early and locally advanced cervical cancers.¹⁰⁴ Lee et al.²⁵ demonstrated that fD^* values positively correlated with DCE-MRI parameter, _estK^trans (estimated volume transfer constant between blood plasma, and the extravascular extracellular space, r = 0.42, P = 0.038). Li et al. recently reported that higher D, f, and V_e values and lower K^trans and K_ep values were observed in cervical carcinoma with high-expression of HIF-1α. DCE-MRI combined with IVIM DWI had higher sensitivity and accuracy than that of DCE-MRI or IVIM DWI for differentiating the high-expression group and the low-expression group of HIF-1α (P = 0.03, 0.02; 0.04, 0.03).¹⁰⁵

Intravoxel incoherent motion has been assessed in cervical cancer treated with chemoradiotherapy (CRT), and was reported to be an early predictor of treatment response.^106,107 The changes in IVIM parameters (D, D^*, f) and ADC before and during CRT were found to be significantly higher in complete remission (CR) than non-CR groups.¹⁰⁸

Others

The f-value was found to correlate with ¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) metabolic parameters in patients with vertebral bone metastases, suggesting its potential for monitoring treatment response.¹⁰⁹ A recent study has shown that histogram analysis of f and D^* may be useful for early response assessment during NAC in osteosarcoma.¹¹⁰ The utility of IVIM parameters has also been demonstrated in differentiating malignant childhood tumor types (Fig. 5).¹¹¹ Other potential clinical applications in oncology include rectal [decrease of ADC, pseudo-diffusion coefficient, and perfusion fraction with poorer tumor differentiation (r = 0.520, P < 0.001; r = 0.447, P = 0.001; r = 0.354, P = 0.010, respectively)],¹¹² esophageal (higher diagnostic performance of ADC_slow than ADC for differentiation of grades of esophageal carcinoma),¹¹³ and lung carcinoma [significantly lower ADC calculated using all b-values, D- and f-values in lung cancer compared with obstructive pulmonary consolidation (P < 0.05)].¹¹⁴ Despite all these promising results, the clinical application of IVIM parameters for the assessment of treatment response has been limited compared with ADC owing to their only moderate repeatability and reproducibility (please also see “Challenges”).¹¹⁵

Fig. 5.

Histologically verified neuroblastoma (grade IV). (a) T₂-weighted and (b) b = 150 images, and (c)–(f) parametric maps [apparent diffusion coefficient (ADC), D, D^*, and f, respectively]. Whole tumor ROI is shown drawn on the parametric maps. The calculated median values of ADC, D, D^*, and f for the drawn ROI were 1155 × 10^–6 mm²/s, 703 × 10^–6 mm²/s, 17,762 × 10^–6 mm²/s, and 23%, respectively. Adapted from Meeus et al.¹¹¹

Challenges

There has been a growing number of publications on IVIM in the past 10 years, presumably due to the IVIM “wake up-call” in 2008,^116,117 reporting significant decrease in D^* and ADC values in cirrhotic patients, which has shed some light on clinical applications of IVIM. Additional IVIM models and fitting methods have been explored in various organs and diseases.¹¹⁸

One needs to keep in mind that the behavior of DW signals is non-Gaussian (monoexponential for tissue diffusion), especially in highly restricted tissues such as cancers, where departure from a Gaussian distribution appears at b-values as low as 600 s/mm² (the highest b-values reported when the initial biexponential model was introduced were approximately 200 s/mm².) Using the standard biexponential IVIM/diffusion model when including high b-values fails to take this effect into account, and the f fraction becomes artificially high (sometimes >40%) due to the presence of residual tissue diffusion effects in the perfusion-related IVIM part of the signal (this artifact would be the largest in tissues with hindered diffusion, i.e., the most malignant areas). In addition, it is well known that fitting data with the biexponential model through the common least-squares model-fitting approach is sensitive to noise effects and outliers.³⁴ Thus, development of a more accurate model is crucial, especially in oncology, and many approaches have been proposed (see “Future Trends”). To take care of non-Gaussian and “noise floor” effects visible when higher b-values are used (approximately 1000 s/mm²), more sophisticated models must be used for data analysis. One popular model is the kurtosis model, where non-Gaussianity in diffusion displacement probability distributions of water molecules in tissues can be measured.^6,119,120 Eventually, more effective IVIM and non-Gaussian DWI models might lead to more accurate handling of IVIM parameters.

Intravoxel incoherent motion parameters (especially D^*) have modest repeatability and reproducibility (Table 2),^121–130 perhaps due to their inherent inflexibility in the standard IVIM model and sensitivity to noise,⁶ and several approaches have been explored to improve their uncertainty. IVIM values are dependent on the acquisition parameters such as b-values,¹¹⁸ TE,¹³¹ and fitting methods,⁸¹ and uniform acquisition of IVIM data and development of robust IVIM fitting are warranted. Many efforts are underway to improve the estimation of IVIM parameters. Removal of motion-contaminated and poorly fitted image data has been proposed to improve their reproducibility.¹³² The uncertainty of IVIM D and f estimates can be reduced by the use of optimized b-value schemes.¹³³ Cardiac gating was found useful in improving the reproducibility of IVIM values in the head and neck.¹²⁵

Table 2.

Repeatability and reproducibility of IVIM parameters in cancers

Authors	Year	Organ	Subjects (n)	Number of b-values	Interval between two scans	Coefficient of variation or (Bland–Altman 95% limits of agreement)
						f	D*	D
Andreou et al.121	2013	Liver	14	8	1 h	(−75.3, 241.0)	(−89.0, 2120.0)	(−20.8, 25.3)
Kakite et al.122	2015	Liver	15	16	5 days	37.3	60.6	19.7
Winfield et al.123	2015	Ovary	31	10	3 days	44	165.1	13.2
Jerome et al.124	2017	Pediatric solid tumors	15	6	1 day	41	35.1	2.5
Kang et al.125	2017	Neck	5	17	10 min	22.1	41.8	8.2
Kang et al.125	2017	Neck	5	Four with cardiac gating	10 min	15.3	29.2	5.7
Sun et al.126	2017	Rectum	26	8	20–30 min	126.3	197.4	24.5
Lecler et al.127	2017	Orbit	22	15	17 min	43	110	14
Pan et al.128	2018	Kidney	25	9–16	1–2 days	26.8–50.5	75.5–101.4	11.8–19.1
Jiang et al.129	2018	Lung	50	10	0.5–1 h	36.5–38.3	68.6–72.6	11.1–11.3
Iima et al.130	2018	Breast	Seven malignant, eight benign	5	Continuous	8.38–16.8	33.4–34.0	–

Future Trends

There remain many challenges in IVIM, both in terms of acquisition and processing. There is little standardized software available to estimate IVIM parameters,⁴⁰ and DWI data often cannot be processed online in picture archiving and communication systems (PACS). Cooperation with vendors is crucial for the development of software that will work in PACS, as well as to generate robust IVIM parameters, and a one-step approach for the optimization and standardization of acquisition protocols and image processing. Efforts are ongoing for optimizing acquisition schemes such as b-values or number of acquisitions within a clinically feasible scanning time.^133–135 Segmented (or two-step, stepwise, over-segmented, or asymptotic) model fitting has been most commonly used to estimate robust diffusion and perfusion parameters.¹¹⁸ Several groups have investigated the utility of Bayesian analysis as an alternative to IVIM model-fitting parameter estimation,^136–138 although parameters might be more biased in hypoperfused tissues with this method. Simplified approaches have attempted by several groups to estimate IVIM parameters without fitting;^72,139 for instance, Sumi et al. estimated IVIM parameters using a geometric method, and Teruel et al. introduced relative-enhanced diffusivity (RED), which is based on the relative increase of the ADC values calculated at a low b-value with respect to the ADC values measured at medium b-values. The optimal b-values for RED have also been investigated, and the authors suggest including b-values of 100 in the breast and 50 in the liver for RED.¹³⁵

Integration of artificial intelligence (AI)/machine learning and IVIM in data acquisition and analysis might be an important approach in the future. A recent study showed a machine-learning algorithm combining MRI-derived data including IVIM as a potential predictive biomarker of treatment outcomes in sinonasal SCCs. Deep neural networks have been explored for accurate and robust IVIM or IVIM and non-Gaussian DWI model fitting to DWI data.^140,141

Other advanced IVIM models are worthy of investigation. Flow-compensated IVIM has been explored, which could be helpful in removing the effect of relatively large vessels on f fraction at the time of acquisition. Diffusion time-dependent IVIM estimates have been demonstrated in the mouse brain; however, significant changes in IVIM values with changing diffusion times have not been identified in xenograft mouse models.¹²⁵

More and more studies have investigated the clinical applications of IVIM in oncology over the past decade. Abundant numbers of approaches and strategies have been extensively explored for optimization of IVIM/diffusion data acquisition and processing. Work remains to improve reproducibility in IVIM parameters and establish the pipeline to analyze IVIM/DWI data online, in cooperation with vendors.