Computed diffusion weighted imaging (cDWI) and voxelwise-computed diffusion weighted imaging (vcDWI) for oncologic liver imaging: A pilot study.

OBJECTIVE: Aim of the study was to evaluate the influence of the selection of measured b-values on the precision of cDWI in the upper abdomen as well as on the lesion contrast of PET-positive liver metastases in cDWI and vcDWI.
METHODS: We performed a retrospective analysis of 10 patients (4 m, 63.5 ± 12.9 y/o) with PET-positive liver metastases examined in 3 T-PET/MRI with b = 100,600,800,1000 and 1500s/mm2. cDWI (cb1000/cb1500) and vcDWI were computed based on following combinations: i) b = 100/600 s/mm2, ii) b = 100/800 s/mm2, iii) b = 100/1000s/mm2, iv) b = 100/600/1000s/mm2 v) all measured b-values. Mean signal intensity (SI) and standard deviation (SD) in the liver, spleen, kidney, bone marrow and in liver lesions were acquired. The coefficient of variation (CV = SD/SI), the differences of SI between measured and calculated high b-value images and the lesion contrast (SI lesion/liver) were computed.
RESULTS: With increasing upper measured b-values, the CV in cDWI and vcDWI decreased (CV in the liver in cb1500: 0.42 with b100/600 s/mm2 and 0.28 with b100/b1000s/mm2) while the differences of measured and calculated b-value images decreased (in the liver in cb1500: 30.7% with b = 100/600 s/mm2, 19.7% with b100/b1000s/mm2). In diffusion-restricted lesions, lesion contrast was at least 1.6 in cb1000 and 1.4 in cb1500, respectively, with an upper measured b-value of b = 800 s/mm2 and 2.1 for vcDWI with an upper measured b-value of b = 1000s/mm2. Overall, the lesion contrast was superior in cb1500 and vcDWI compared to cb1000 (15% and 11%, respectively).
CONCLUSION: Measuring higher upper b-values seems to lead to more precise computed high b-value images and a decrease of CV. vcDWI provides a comparable lesion contrast to b = 1500s/mm2 and offers additionally the reduction of T2 shine-through effects. For vcDWI, measuring b = 1000s/mm2 as upper b-value seems to be necessary to guarantee good lesion visibility in the liver based on our preliminary results.

Introduction

Diffusion-weighted imaging (DWI) has become one of the most widely used functional imaging techniques in magnetic resonance imaging (MRI). Within a few minutes, DWI is able to provide tissue information on a molecular scale [1]. Technical developments such as the introduction of single shot echo planar imaging (EPI) and parallel imaging improved image quality and allowed for the application of DWI in extracranial regions [2,3]. While malignant tumors usually show different tissue characteristics to the tissue they arise from (such as higher cellularity or the integrity of cell membranes), DWI nowadays plays a pivotal role in oncologic abdominal imaging [4,5]. Technically, DWI is based on a T2-weighted spin-echo EPI sequence modified by diffusion-sensitizing paired gradients [6]. The sensitivity can be varied by the time interval between the gradients, the duration and the amplitude of the applied gradients which is subsumed under the term “b-value”. As DWI is based on a T2-weighted sequence, the signal intensity in b-value images does not only depend on the diffusivity of water molecules but also on the T2 relaxation properties of the investigated tissue. This is known as T2 shine-through effect and might result in misleading interpretations. It has been shown that high b-value images of up to b = 1000–1500 s/mm2 can improve tumor detection in selected anatomic sites [7,8]. As acquiring high b-value images can time consuming and more prone to image artifacts as compared to lower b-value images, they have not been implemented in daily routine of whole-body imaging so far [9,10]. With the aim to improve image quality, Blackledge et al. proposed an approach to compute high b-value images based on lower measured b-value images: computed DWI (cDWI) [11]. This has mostly been evaluated in the prostate [9,10,12,13]. However, the precision of computed high b-value images and the dependence on measured lower b-value images have not been investigated yet in the upper abdomen. Recently, Gatidis et al. proposed a new voxelwise computed DWI (vcDWI) technique to further improve the visibility of diffusion restricted lesions [14]. In contrast to the method by Blackledge et al., it computes the presented b-value image for each voxel in dependence on its apparent diffusion coefficient (ADC) and calculates its respective intensity value. Thereby, voxels with low ADC are presented with signal intensity of low b-values and vice versa. This should improve the contrast of diffusion restricted lesions and reduce the T2 shine-through effect.

The aim of our study was twofold: First, to investigate the influence of the selection of measured b-value images on the precision of computed high b-value images (cDWI) in the upper abdomen. Second, to evaluate quantitative image features of cDWI and vcDWI in organs and metastatic liver lesions in dependence on the measured b-value images used for the computation.

Material and methods

Patient cohort

The data of 10 consecutive patients (4 male, mean age 63.5 ± 12.9 years) with PET-positive liver metastases and a PET/MRI protocol including DWI with high b-value images (up to b = 1500s/mm2) were retrospectively evaluated. The local ethics committee waived informed consent for the retrospective evaluation of the data. The oncologic diseases were distributed as follows: Melanoma (n = 5), neuroendocrine tumor (n = 3), adenocarcinoma of the small bowel (n = 1), breast cancer (n = 1). Metastatic involvement of the liver was histology-proven in four patients. In six patients, follow-up examinations revealed progressive metastatic disease of the liver (n = 5) or response under therapy (n = 1).

PET/MRI protocol

All patients were examined in a simultaneous 3 T PET/MRI-scanner (Biograph mMR, Siemens Healthcare GmbH, Erlangen, Germany). A 2D single-shot spin-echo EPI sequence in 3-scan-trace mode with monopolar diffusion gradients and five different b-values (b = 100, 600, 800, 1000 and 1500s/mm2) was applied in the upper abdomen of each patient. Sequence parameters are given in Table 1. Additionally, a navigator-triggered T2-weighted 3D fast-spin-echo sequence (T2-TSE) was performed. Other sequences were chosen depending on clinical indication. Depending on the disease, 18F-FDG (melanoma, adenocarcinoma, breast cancer) or 68Ga-DOMITATE (neuroendocrine tumor) was used as PET-tracer.

Table 1

Sequence parameters of DWI. Examinations were performed in free breathing.

	DWI
Echo Time	65 ms
Repetition Time	7300 ms
Matrix Size	168 × 192
Pixel Size	2.2 x 2.2 mm
Slice Thickness	5 mm
Number of slices	34
Pixel Bandwidth	1736 Hz/pixel
Acquisition Time	2.5 min
Field of view	350 × 400
Number of averages	5
Acquisition plane	axial
b-values (s/mm2)	100, 600, 800, 1000, 1500
Fat suppression method	Spectral attenuated inversion recovery fat suppression

The computation of cDWI/vcDWI

While in the cDWI method the signal intensity of each image voxel is calculated for a predefined constant b-value [11], in the vcDWI method the chosen b-value for each voxel varies dependent on its ADC-value: Sb(x) = S0(x) * exp(-ADC(x) * (k*ADC(x)-b0), where Sb(x) is the calculated signal intensity of voxel x for a b-value (k*ADC(x)), S0(x) its measured signal intensity with b-value = 0 s/mm2 (b0), ADC(x) is its ADC value and k = 106 s2/mm4. The ADC-dependent choice of the voxelwise-computed b-value can thus increase the contrast between diffusion-restricted and unrestricted tissues because signal intensities of voxels with low ADC are computed at lower b-values while voxels with high ADC at higher b-values [14].

VcDWI as well as cDWI images for b = 1000 and b = 1500s/mm2 (cb1000 and cb1500, respectively) were calculated based on a mono-exponential model from five different combinations of b-value images: i) b = 100 and 600 s/mm2 (b100/600), ii) b = 100 and 800 s/mm2 (b100/800), iii) b = 100 and 1000s/mm2 (b100/1000), iv) b = 100, 600 and 1000s/mm2 (b100/600/1000) v) all measured b-value images (b_all). The calculation of cDWI and vcDWI was carried out as described in Blackledge et al. [11] and the recently published work by Gatidis et al. [14] using MATLAB (The MathWorks Inc, Natick, MA).

Image analysis

To avoid partial volume effects, only lesions with a diameter of >1 cm were included to the quantitative evaluation. Lesions were rated as PET-positive if the focal tracer uptake exceeded the regional uptake of physiological liver tissue. For anatomical correlation, the T2-TSE images were rigidly registered to the b = 800 images. Regions of interest (ROIs) were drawn freehand in the b = 800 images in up to three PET-positive lesions of the liver by F.S. (6 years of experience in MRI, 4 years of experience in hybrid imaging). Furthermore, circular ROIs were set in visually not affected parenchyma of the right and left liver lobe, the spleen, the right or left kidney, the psoas muscle, the second lumbar vertebra (bone marrow) and the image background. Care was taken to avoid image artifacts and borders of organs and lesions. ROIs in physiological tissue and background had a target diameter of 1.5 cm in the liver and 1 cm in other organs. The ROIs were copied to the different measured and computed b-value images. Those steps were performed using PMod (PMOD Technologies Ltd, Zurich, Switzerland). Mean signal intensities (SI) were measured in each ROI. Standard deviations (SD) were acquired in tissue ROIs in cDWI and vcDWI. SI of the liver was defined as the mean value of SI in the right and left liver lobe. Measured b = 1000s/mm2 and b = 1500s/mm2 images are abbreviated as “mb1000” or “mb1500” in the following, respectively.

The following parameters were calculated:

The relative signal difference of physiological tissue ROIs between the measured and calculated (cDWI) b-value images:Relative differences = abs (SI calculated b-value image – SI measured b-value image) / SI measured b-value image * 100.

The contrast (signal intensity ratio) of the lesion ROIs to the surrounding liver tissue (liver ROI) for the different measured b-value images as well as for cDWI and vcDWI:Lesion contrast = lesion SI / liver SI.

The coefficient of variation within the different physiological tissue ROIs for the different measured b-value images as well as for cDWI and vcDWI as an indicator of the image noise:CV = SD / SI.

Results

In total, 20 liver metastases (in 10 patients) with diameters ranging from 1.0 to 4.8 cm were evaluated. The SI, the relative differences between measured and calculated b-value images, the CV and the lesion contrast are shown in Fig. 1, Fig. 2, Fig. 3, Fig. 4 respectively.

Fig. 1

Whisker box plots including median, first and third quartile, maximum and minimum of lesion contrast of measured and calculated b-values as well as in vcDWI. The mean lesion contrasts of cb1500 and vcDWI seem to be on a comparable level. However, highest variances in computed b-value images as well as in vcDWI are seen with b = 100/600, resulting in values partially close to zero.

Fig. 2

Mean signal intensities of measured and computed b-values as well as vcDWI in lesions and organs. Note the superior signal intensity in lesions in vcDWI compared to measured and calculated b1500. Physiological spleen tissue also shows high signal intensity caused by high cellularity.

Fig. 3

Coefficient of variation in physiological tissue and image background. Taking higher b-value images for the calculation of cb1000 and cb1500 decreases the covariance. Overall, vcDWI shows higher CV in physiological tissue; the high CV in vcDWI in kidneys and background are mainly caused by the very low signal intensity in those ROIs (see Fig. 2).

Fig. 4

Relative differences of the mean signal intensities between measured and computed b-value images in %. A slight trend towards decreasing deviations with increasing upper b-values is present while b = 100/1000 provides overall best results for most organs in cb1500 from the b-value combinations tested.

The SI in physiological tissue and lesions in mb1500 images were about 39% lower than in mb1000 images. For cDWI, SI in physiological tissues showed to be relatively stable and rather independent on the measured b-values used for calculation (max. 23% difference for cb1000 in lesions between b100/600 and b100/1000). VcDWI provided SI approximately on the level of b = 1000 images with highest SI for b100/1000 calculations, especially in lesions (Fig. 2). In the spleen, as an organ with inherently low diffusivity, vcDWI provided higher SI as compared to all other measured and calculated images. However, an overall higher dependence of lesion SI on b-value selection – especially for lesion ROIs – was found in vcDWI.

While cb1000 images showed only a slight underestimation of mean signal intensities as compared to mb1000 images, a pronounced underestimation was seen in the cb1500 images on average (maximum of 15.7 and 28.3%, respectively, Fig. 2, Fig. 4). With increasing upper measured b-value images, the relative differences between measured and calculated b-value images showed a decreasing trend.

Overall, the CVs of cb1500 images were higher than in cb1000 images (Fig. 3). The CVs in cDWI and vcDWI showed a decreasing trend with increasing upper measured b-value images. However, the CVs in high b-values images calculated with b100/600/1000 were higher than for b100/1000 in the kidney and liver. In vcDWI, the CV was higher in the liver, the kidneys and the muscle as in computed b-value images while it was on a comparable level in the spleen and the bone marrow (CV in the liver when computation was based on b100/800: cb1000, 0.23; cb1500, 0.35; vcDWI, 0.48). In the image background, the CV in vcDWI was lower than in the cb1500 images.

The lesion contrast was slightly superior in cb1500- and vcDWI-images as compared to cb1000-images (maximum 15% and 11%, respectively; Fig. 1). On the other hand, cb1500 and vcDWI images also provided higher variances in lesion contrast; the highest ranges of lesion contrast were found in cDWI and vcDWI when based on b100/600 (min/max: cb1000, 0.4/10.1; cb1500, 0.2/13.1; vcDWI, 0/15.6).

Fig. 5 gives an example of a patient with liver metastases of melanoma to demonstrate the image impression and lesion contrast of the different measured and calculated b-value images as well as the vcDWI. The smaller lesion, marked with a dotted arrow in the Figure, is smaller than 1 cm in diameter and was therefore not included to the quantitative evaluation.

Fig. 5

Example of a 60 y/o male patient with melanoma and metastatic liver lesions. The larger lesion (white arrow) can clearly be seen in vcDWI as well as in computed and measured b-values (cb and mb, respectively); note the differences in lesion-to-liver contrast and image noise. One small lesion (dotted arrow) is visible in measured b-values (mb1000 and mb1500) but not in the calculated b-values (cb1000 and cb1500) based on b = 100/800. In vcDWI, this small lesion is clearly visible if the calculation is based on b = 100/1000; In contrast, the lesion is masked if the calculation is based on b = 100/800.

Discussion

In this study, we investigated the influence of the selection of measured b-value images on the precision, the image characteristics and the lesion contrast of cDWI and vcDWI approaches in the upper abdomen.

Diffusion-weighted imaging is based on a T2-weighted sequence. Therefore, the signal intensities in images are influenced by both the T2 properties and the diffusion restriction of the investigated tissue. With increasing b-values, the diffusion-weighting increases and thus the SI decreases (especially for less diffusion-restricted tissues). As expected, the mb1500 images consecutively showed lower signal intensities than mb1000 images in our study. The mean signal intensities of computed b-value images seem to be rather stable and independent on the used measured b-value images for calculation. While the computation of cb1000 images was relatively precise, a pronounced underestimation of cb1500 images on average was observed. The computation of high b-value images in our study is based on a mono-exponential fit which is a model assuming free Gaussian diffusion as a normal distribution of tissue diffusivities. However, as shown in several studies, this is only a simplified model which does not reflect the true diffusion properties of human tissue [[15], [16], [17]]. Several approaches have been proposed trying to take these effects into account like bi-exponential models or diffusion kurtosis imaging [18,19]. However, they have not been implemented into daily routine of abdominal DWI so far. The multi-exponential dependence of the diffusion signal on b-value images (or diffusion-weighting) is the main reason for the difference between measured and computed b-value images. If a higher upper b-value is close to the value to be computed, the deviation to the measured value in a mono-exponential fit is expected to be low. Overall, a trend towards lower deviations with higher upper measured b-value images was observed in our study which is in concordance to previously reported results [20]. Using the b = 600 s/mm2 value to the calculation of cb1500 decreased the precision of calculation in most tissue types in organs with higher perfusion content (liver and kidney). Therefore, from a quantitative point of view b = 600 s/mm2 seems not to be valid for calculation of high b-value images in those organs.

The differences of measured and calculated b-value images were highest in the kidneys. A reason might be that the kidneys are organs with a high portion of capillary perfusion and an anisotropic diffusion; therefore it is likely that the b100 value is strongly influenced by perfusion effects [21]. To reduce the influence of these perfusion effects on DWI, previously conducted studies suggest b-values of at least b = 800 s/mm2 [22]. This can also be seen in our study in a noticeable drop of computed-to-measured deviations from b = 100/600 to b = 100/800 which is synonymous with a more precise ADC measurement.

As the SI in calculated b-value images is relatively stable, the decreasing trend seen in CVs is mainly caused by decreasing standard deviations. As standard deviation represents the inhomogeneity of signal intensities within a ROI, the reason for this behaviour might be found in the more precise voxelwise computation of high b-value images with higher upper b-value images. Therefore, the course of CVs resembles the course of the relative differences between measured and calculated b-value images. For vcDWI, the b-value of each voxel is calculated based on the ADC-measurement. Therefore, deviations in the ADC-map lead to even more pronounced deviations in the vcDWI-images as compared to the calculated high b-value images. This is represented in the higher CVs in physiological tissue. In contrast, the CVs in the image background in vcDWI is relatively low.

In concordance to the behaviour of CV and differences in signal intensities, highest variances in lesion contrast of cb1000, cb1500 and vcDWI occurred if the calculation was based on b100/600. This led to a lesion contrast of <1 in calculated b-value images in two lesions and in vcDWI in five lesions out of 20. This is crucial as it might result in misinterpretations in oncologic reading. In one of the patients, the investigated 68Ga-PET-positive lesion showed low signal intensities and a lesion contrast below 1 in most calculated b-value images, in mb1500 and even close to zero in vcDWI, respectively. This example is demonstrated in Fig. 6. This patient suffered from a metastasized neuroendocrine tumor and had undergone systemic therapy. As a possible response to therapy it can be seen that metastases change their tissue properties caused by cell swelling, tumor lysis and necrosis [[23], [24], [25], [26]]. Therefore, the diffusivity can increase and the signal intensity in high b-values and vcDWI therefore decrease. The very low lesion contrast in this case is therefore not to be interpreted as an incorrect computation of b-value images or a drawback of vcDWI but as a therapy response and a general limitation in the detection of liver metastases in DWI as DWI is merely able to detect diffusion restricted lesions. Thus, in this case the lesion contrast in measured b-value images was most probably not due to differences in tissue diffusion restriction but because of differences in T2 relaxation times. Since the vcDWI method reduces such effects, this result in the observed loss in lesion liver contrast. This should be kept in mind when using vcDWI. Besides this case, the lesion contrast in computed high b-value images was at least 1.6 (cb1000) and 1.4 (cb1500), respectively, when an upper measured b-value of at least 800 s/mm2 was chosen. For vcDWI, an upper b-value of 1000s/mm2 should be preferred to maximize lesion SI and minimize CV resulting in a reliable lesion contrast of at least 2.1 in this study (besides the previously described case).

Fig. 6

Example of a 71 y/o female patient with 68Ga-PET-positive liver metastases of a neuroendocrine tumor. The patient had undergone a partial liver resection and chemotherapy. The lesion shows a distinct PET tracer uptake (top row, left hand side: PET overlaid with T2 TSE) which stands for a high expression of the somatostatin receptor (SSR). The lesion does not provide a considerable diffusion restriction which is a common finding in metastases of NET under therapy. Thus, the lesion is not distinguishable in b = 1500 images (mb1500/cb1500) or vcDWI. However, in b = 1000 images (mb1000/cb1000), the lesion is visible which can be explained by a T2-shine-through effect. All calculated b-values as well as the vcDWI were based on b = 100/800.

Our study has several limitations. Due to the small population size with different types of liver metastases we were only able to detect general trends and a potential area of application. Statistically significant differences could not be extracted since the signal in b-images is not a quantitative parameter and thus varies strongly between patients. On the other hand, our cohort represents a broad spectrum of different kinds of metastatic liver lesions showing the general usability of the methods. The effectiveness of the different methods on specific diagnostic questions has to be investigated in future on dedicated clinical trials. The minimum diameter of included lesions was set to 1 cm to reduce partial volume effects for quantitative analysis. Therefore, the detectability of small lesions was not evaluated. Liver tissue without PET-positive lesions and other visible pathologies was defined as physiological; other diffuse inhomogeneities in liver parenchyma like small areas of fibrosis or steatosis which might have influence on diffusivity were not respected. Finally, the monoexponential fit is a simplified model which does not consider perfusion effects or other diffusion effects at high b-values, however, it is the most widely used approach in DWI allowing to quantify diffusion with only two b-value images.

In conclusion, we could show that the precision and the image characteristics of cDWI and vcDWI can be dependent on the selection of measured b-values the calculations are based on and in individual cases, the selection can have influence on the visibility of metastatic liver lesions. Higher measured upper b-values lead to more precise calculations of computed high b-value images with lower CVs. Based on our preliminary results, b = 800 s/mm2 as upper measured b-value seems therefore to be necessary to avoid misinterpretations in oncologic readings in the upper abdomen with cb1000 and cb1500, respectively. Adding additional lower b-value images to the monoexponential calculation seems not to improve the reliability of computed high b-value images, especially in the kidneys. vcDWI provides a comparable lesion contrast to b = 1500s/mm2 images with higher signal intensities of lesions and offers additional features such as the reduction of T2 shine-through effects which could help to detect diffusion restricted liver lesions. However, caused by the underlying algorithm vcDWI is more susceptible to errors in the ADC-map; measuring b = 1000s/mm2 as upper b-value seems therefore to be necessary to guarantee good lesion visibility in the liver.

Conflict of interest

The authors declare that they have no conflicts of interest.

Declarations of interest: none