Breast Cancer Conspicuity on Computed versus Acquired High b-Value Diffusion- Weighted MRI

Abstract

Rationale and Objectives:

On unenhanced diffusion-weighted imaging (DWI), computing or synthesizing high b-value images from lower b-value acquisitions can enhance breast cancer visibility. This study aimed to evaluate relative lesion conspicuity on computed versus acquired diffusion-weighted images and investigate clinical characteristics influencing optimal b-values.

Materials and Methods:

Women with newly diagnosed breast cancer were prospectively enrolled and underwent 3T breast MRI with DWI. Lesion contrast-to-noise ratio (CNR) was measured across a range of b-values (0–2500 s/mm²) for computed and acquired DWI. Three readers independently compared lesion visibility between computed and acquired DWI and selected the optimal b-value. Computed versus acquired DWI was compared quantitatively based on CNR by paired t-test and qualitatively based on reader preference using a sign test. Optimal b-values by qualitative and quantitative assessment were compared by paired t-test, and associations with clinical characteristics were assessed by Wilcoxon rank sum test.

Results:

The study included 30 women (median age, 48 years); 28 with invasive carcinoma, 2 DCIS. Lesion CNR was higher on acquired versus computed images (p=0.018), while lesion visibility by reader assessment was not different (p=0.36). Optimal b-values selected by readers (mean, b=1411±383 s/mm²) were slightly higher than those based on peak CNR (b=1233±463 s/mm², p=0.023), and were higher for younger (≤50 years) versus older women (p=0.002) and dense versus non-dense breasts (p=0.015).

Conclusion:

Lesion CNR on computed high b-value images was slightly reduced versus acquired images, but our study suggests that this did not significantly impact lesion visibility. Computing high b-value images offers extra flexibility to adjust b-value during interpretation.

INTRODUCTION

Dynamic contrast-enhanced magnetic resonance imaging (DCE MRI) is both a sensitive and specific tool for breast cancer detection. DCE MRI has proven to be valuable for supplemental screening for women at high risk of developing breast cancer and for preoperative evaluation of extent of disease (1–5). However, the reliance of DCE MRI on gadolinium-based contrast agents requiring administration and monitoring by trained medical staff contributes to both long examination times and high costs. Additionally, recent studies have shown that gadolinium deposits remain in the body long after injection, posing concerns about potential patient safety (6). In response to these concerns, there is a strong interest to identify a safe, quick, and cost-effective alternative to DCE MRI for supplemental breast imaging.

Diffusion weighted imaging (DWI) has the potential to avoid some of the disadvantages of DCE MRI and provide a quick, unenhanced tool for breast cancer detection and characterization (7). DWI measures the diffusion of water in tissue using motion sensitizing gradients, giving information about the microscopic cellular environment. In DWI, the signal intensity is inversely proportional to the average mobility of water in a voxel (8). DWI allows for imaging of breast cancer without the use of an exogenous contrast agent in part because the higher cell density that characterizes malignant tumors hinders diffusion, making the tumors appear hyperintense in relation to normal breast tissue.

The diffusion sensitization factor or b-value refers to the strength and timing of the magnetic gradient applied during DWI acquisition (8). Increasing the b-value during image acquisition can increase breast lesion contrast due to higher diffusivity in normal fibroglandular tissue (FGT) versus malignancy (9, 10). As the b-value increases, the signal intensity of the normal FGT decreases at a rate faster than that of tumors, making malignancies more conspicuous. However, optimal b-values for lesion detection are not one-size-fits-all and likely vary with both patient and tumor characteristics (e.g., breast density, lesion size, morphology, etc.); it is not clear what generalizable high b-value should be used. Furthermore, acquiring images at these high b-values increases image distortions due to susceptibility effects and eddy- currents, degrades signal-to-noise ratio (SNR), and lengthens scan times (11). An alternative approach is to compute high b-value images from images acquired at lower b-values (referred to as ‘computed’ or ‘synthesized’ DWI) rather than acquiring them directly (12), which is now provided by some MRI vendors as a console software tool. This approach has been a growing area of investigation in breast imaging (7, 9, 13–17). However, it is also recognized that synthesized DWI boosts contrast at the expense of accurate microstructural characterization (18), and it is not clear if synthesized DWI offers a clear advantage over acquired DWI.

Therefore, the purpose of this study was to evaluate the relative conspicuity of breast cancer on synthesized versus acquired DWI across a wide range of b-values (0 – 2500 s/mm²), and to investigate clinical characteristics influencing optimal b-values, in a consecutive cohort of women with newly diagnosed breast cancer.

MATERIALS and METHODS

Patients

This study was approved by our Institutional Review Board and was Health Insurance Portability and Accountability Act compliant. From July 2018 to June 2019, consecutive patients ≥ 18 years old with biopsy-confirmed breast cancer of any size or type were prospectively consented to participate in the study. Enrolled patients agreed to undergo an additional research DWI sequence appended to their clinical pretreatment breast MRI exam. Patients were excluded for concurrent neoadjuvant chemotherapy.

MRI Acquisition

MRI acquisitions were performed on a 3 tesla clinical scanner (Achieva, Philips Healthcare, Best, Netherlands) using a 16-channel breast coil (MammoTrak, Philips Healthcare).

Phantom Scans:

Initial DWI experiments were performed using a diffusion breast phantom (QalibreMD, Boulder, CO). Phantom DWI scans were acquired with TR/TE = 3500/79.7 ms, number of signal averages (NSA) = 2, and b-values 0, 100, 800, 1200, 1500, 1800, 2200, 2500 s/mm². Imaging was performed with SPAIR and gradient reversal fat suppression, FOV = 360 × 360 mm², 30 slices, pixel size = 1.8 × 1.8 × 4 mm³, in plane SENSE factor = 3, and MB-SENSE factor = 2. Two scans were performed to assess the influence of high b averaging, which is commonly used to maintain SNR by obtaining an increasing number of signal averages (NSA) with increasing b-value. The first scan was acquired with high b averaging = on (where NSA = 2 for b-values 0–499 s/mm², NSA = 4 for b-values 500–999 s/mm², and NSA = 6 for b-values 1000–2500 s/mm²) with a scan time of 6:29 min. The second scan was acquired with high b averaging = off (where NSA = 2 for all b-values) and a scan time of 2:29 min.

Patient Scans:

Clinical breast MRI examinations were performed in accordance with ACR accreditation guidelines (19), and included axial DWI, T2-weighted fast spin-echo, T1-weighted non fat-suppressed, and T1-weighted fat-suppressed DCE MRI sequences with one precontrast and three postcontrast acquisitions. Research DWI sequences for the study were appended to the end of the clinical protocol. Research DWI scans were acquired with TR/TE = 3500/79.9 ms and b = 0, 100, 800, 1500, 2500 s/mm². To maximize SNR, high b-averaging was used; NSA = 2 for b = 0 and 100 s/mm², NSA = 4 for b = 800 s/mm², and NSA = 6 for b = 1500 and 2500 s/mm². Scans were acquired with SPAIR and gradient reversal fat suppression, FOV = 360 × 360 mm², 30 slices, pixel size = 1.8 × 1.8 × 4 mm³, in plane SENSE factor = 3, and MultiBand-SENSE factor = 2, with scan time of 3:33 min.

Post Processing

All diffusion-weighted images were first registered to correct for eddy current and patient motion effects using a diffusion registration tool (Philips Healthcare, Best, the Netherlands) to align diffusion-weighted images at each b-value and slice position to the corresponding b=0 s/mm² reference image. Each of the patient and phantom scans was processed offline using custom software developed in MATLAB (MathWorks, Natick, MA). First, the reference b = 100 s/mm² image was thresholded to minimize noise and unsuppressed fat signal. Any pixels present on the b = 100 s/mm² image with values below the user-defined threshold were set to zero.

Next, the thresholded b = 100 s/mm² images and b = 800 s/mm² images were fit on a voxelwise basis to a monoexponential decay model to calculate the apparent diffusion coefficient (ADC) map (20):

$$ADC=-\frac{1}{\Delta b}\ln \left(\frac{S_{800}}{S_{100}}\right)$$

where S₈₀₀ is the signal intensity of the image acquired at b = 800 s/mm², S₁₀₀ is the signal intensity of the image acquired at b = 100 s/mm², and Δb is the difference in b-value between the two images (700 s/mm²) (8). The b = 100 s/mm² image was chosen as a reference image rather than the b = 0 s/mm² to minimize contributions from perfusion. The 800 s/mm² image was chosen for creating ADC maps as it is a commonly acquired b-value for breast DWI. High b- value diffusion weighted images were then synthesized by extrapolation from the known ADC values using the equation:

$$S_b=S_{100}{e}^{-\Delta b\cdot ADC}$$

where S_b is the signal intensity of the computed image at any b-value. High b-value images were computed for b-values over the range 200–2500 s/mm² at increments of 100 s/mm².

Assessment of Lesion Conspicuity

Quantitative Assessment

Lesion conspicuity was evaluated quantitatively through contrast-to-noise ratio (CNR) of the lesion to the surrounding or closest normal FGT. Regions of interest (ROIs) used to compute CNR were defined using a semi-automatic segmenting tool in MATLAB. ROIs were drawn by a research scientist (ANONYMIZED-FOR-REVIEW) in consultation with a radiologist (ANONYMIZED-FOR-REVIEW) with 9 years of experience in clinical breast imaging. ROIs for each patient were drawn on a single slice containing the largest measurable portion of the lesion using the DCE image for reference. Lesions and normal FGT were segmented on the b = 800 s/mm² and b = 100 s/mm² images (common to both acquired and computed datasets), respectively (Figure 1). The lower b-value image was used for normal tissue segmentation to maximize normal tissue visibility. Normal FGT measures were taken in the ipsilateral breast on the same slice as the lesion when FGT was present. When no FGT was present in the ipsilateral breast in the lesion slice, normal tissue measures were taken from the contralateral breast in the lesion slice. The lesion and normal FGT ROIs drawn on the b = 800 and b = 100 s/mm² images, respectively, were propagated to all other computed and acquired images at the corresponding slice level.

Figure 1.

Example of lesion and normal FGT segmentations performed for CNR calculations. Regions of interest (ROIs) were defined on the DWI slice with the largest measurable lesion area. Shown are a) dynamic contrast enhanced image used for reference depicting enhancing malignancy (arrow), b) lesion ROI (red contour) defined on the b = 800 s/mm² image, c) normal FGT ROI (yellow contour) defined on the b = 100 s/mm² image, and both ROIs propagated to the b = 1500 s/mm² diffusion-weighted image for CNR calculation (CNR = 2.3).

For measuring CNR in phantom scans, a vial of 40% polyvinylpyrrolidone (PVP) in water (ADC = 0.676 × 10⁻³ mm²/sec at 20° C) was used as lesion mimic. A vial of water (ADC = 2.025 × 10⁻³ mm²/sec at 20° C) was used as normal FGT mimic (21). Standardized rectangular ROIs were drawn to segment lesion and normal tissue mimic vials on the b = 800 s/mm² and b = 100 s/mm² images, respectively (Figure 2), and then propagated to all other computed and acquired b-value images.

Figure 2.

Process for CNR calculations in phantom scans. Segmentations were performed on DWI for a) a vial of pure water (ADC = 2.025 × 10⁻³ mm²/sec at 20°C) as normal tissue mimic on b = 100 s/mm² image, and b) a vial of 40% polyvinylpyrrolidone (PVP) in water (ADC = 0.676 × 10⁻³ mm²/sec at 20°C) as lesion mimic on b = 800 s/mm² image. The ROIs (red outlines) were propagated to corresponding diffusion-weighted images at each b-value for CNR calculations.

Using the lesion and normal tissue ROIs, CNR was then calculated for each b-value using the equation:

$$CNR=\left({\mu }_{\text{lesion}}-{\mu }_{\text{tissue}}\right)/\sqrt{{\sigma }_{\text{lesion}}^2+{\sigma }_{\text{tissue}}^2}$$

where μ_lesion is the mean signal intensity of the lesion, μ_tissue is the mean signal intensity of the normal FGT, σ_lesion is the standard deviation of the signal intensity of the lesion, and σ_tissue is the standard deviation of the signal intensity of the normal tissue.

Qualitative Assessment

A reader study was performed to evaluate lesion conspicuity on computed and acquired images and across b-values. Three fellowship-trained breast radiologists (ANONYMIZED-FOR- REVIEW with 9, 8, and 3 years of experience in clinical breast MR, respectively) independently read the images. For each of the 30 patient cases, readers first reviewed computed and acquired images at b = 1500 s/mm² side-by-side and specified whether the lesion was more conspicuous on one versus the other or if they were equal. Readers were blinded to which image was computed and which was acquired. Readers were given the lesion slice and location along with the corresponding DCE MRI images for reference.

Next, readers determined the b-value which exhibited the highest lesion conspicuity for each patient case. To find the optimal b-value, the reader used a custom software tool to view images of varying b-values. The software tool allowed users to scroll through images of b- values 0–2500 s/mm² at increments of 50 s/mm².

Statistical Analysis

The paired t-test was used to compare CNRs from computed and acquired images and to compare optimal b-values selected by readers and by maximizing CNR. Readers preferences of lesion visibility on computed and acquired images at b = 1500 s/mm² were compared using the sign test. The non-independence of the multiple reads per image was accounted using permutation tests and the bootstrap with resampling performed by patient. Inter-reader agreement in selected optimal b-values was summarized using the intraclass correlation coefficient (ICC). The ICC was also used to summarize agreement between reader-selected optimal b-values (averaged across 3 readers) and optimal b-values selected by maximizing CNR. Associations of reader-selected optimal b-values with clinical factors (age, breast density, lesion size, lesion type, histology, and cancer grade) was assessed using the Wilcoxon rank sum test. All statistical calculations were conducted with the statistical computing language R (version 3.5.1; R Foundation for Statistical Computing, Vienna, Austria).

The study protocol was approved by the Institutional Review Board of the (institution ANONYMIZED FOR REVIEW).

RESULTS

Thirty-eight patients scheduled to undergo a preoperative breast MRI were prospectively consented to participate in this study. Of those, two patients were excluded due to concurrent neoadjuvant chemotherapy and six patients were excluded because biopsy-related artifacts limited lesion evaluation. Of the thirty patients included in the final dataset the median age was 48 years (range: 28 – 83 years) and the median lesion size was 20 mm (range: 8 – 124 mm), Table 1.

Table 1.

Patient and tumor characteristics.

Variable	Value
Age
≤ 50 years	20 (66.7)
> 50 years	10 (33.3)
Breast density
Dense	25 (83.3)
Not dense	5 (16.7)
Size
≤ 2 cm	17 (56.7)
> 2 cm	13 (43.3)
Lesion type
Mass	25 (83.3)
NME	5 (16.7)
Histology
IDC	26 (86.7)
ILC	2 (6.7)
DCIS	2 (6.7)
Nottingham grade
1	9 (30.0)
2	4 (13.3)
3	15 (50.0)
Not available	2 (6.7)

Values are no. (%).

Abbreviations: NME = non-mass enhancement, IDC = invasive ductal carcinoma, ILC = invasive lobular carcinoma, DCIS = ductal carcinoma in situ

Computed versus Acquired DWI

Phantom Scans

Two phantom DWI scans were performed to assess the influence of b-value and high b averaging on CNR and to compare computed and acquired DW images. Lesion-to-normal tissue (mimic) CNR measures increased with b-value for both acquired and computed images, Figure 3. Without high b averaging, CNRs of acquired and computed images were similar across all b- values. However, when high b averaging was used, CNR measures were slightly higher for acquired versus computed images at higher b-values (b > 800 s/mm²). Both acquired and computed images had higher CNR when high b averaging was used.

Figure 3.

Contrast-to-noise ratio in the phantom as a function of b-value in both acquired and computed diffusion-weighted images. CNR increases with increasing b-value and by using high b averaging for both acquired and computed images.

Patient Scans

Quantitative Assessment:

To maximize SNR, high b averaging was used for all in vivo scans. Overall, mean lesion CNR increased with increasing b-value, peaking near b = 1200 s/mm² for computed images and 1500 s/mm² for acquired images (Figure 4). CNR began decreasing at b-values greater than 1500 s/mm². Similarly to the phantom observations, higher mean CNR was observed in acquired vs. computed scans at higher b-values (b > 800 s/mm²), both at b = 1500 s/mm² (mean: 2.6 vs. 2.4, difference: 0.2, 95% CI: 0.04 to 0.4, p = 0.018) and b = 2500 s/mm² (mean: 2.3 vs. 1.9, difference: 0.4, 95% CI: 0.2 to 0.6, p = 0.003). Example computed high-b value images in a study patient with invasive ductal carcinoma are shown in Figure 5.

Figure 4.

Mean in vivo lesion contrast-to-noise ratio (CNR) as a function of b-value for computed and acquired diffusion-weighted images in 30 patients with breast cancer. CNR generally increased with increasing b-value, peaking near 1200 s/mm² for computed and 1500 s/mm² for acquired images. CNR was significantly higher for acquired vs. computed images at both b = 1500 s/mm² and b = 2500 s/mm² (p = 0.018 and p = 0.003, respectively, paired t-test).

Figure 5.

Example images in a 42-year-old patient with a 20 mm invasive ductal carcinoma and extremely dense breasts. Shown are the reference DCE MRI image depicting the malignant mass (arrow), acquired (‘Acq’) DW images at b = 0, 100, 800 s/mm², and computed (‘Comp’) DW images at b = 1000, 1200, 1400, 1600, 1800, 2000 s/mm².

Qualitative Assessment:

Across the three reader assessments of the 30 patients (90 reads total), lesion visibility at b = 1500 s/mm² was rated as similar for computed and acquired images in 65 of 90 (72%). The acquired images were preferred in 16 (18%) and the computed images were preferred in 9 (10%, p=0.36 for the difference). Overall, the computed images were rated as at least as good as the acquired in 74 of 90 (82%, 95% CI: 71–92%). This percentage ranged from 73% to 90% across the 3 readers (Figure 6).

Figure 6.

Stacked bar chart of reader preferences for acquired versus computed images based on lesion visibility at b = 1500 s/mm².

Optimal b-Values

Across individual patients, the optimal b-value for lesion conspicuity varied widely based on both quantitative and qualitative assessment. Two example cases are shown in Figure 7. Quantitative assessment showed the optimal b-value for peak lesion CNR on computed images ranged from b = 400 to 2500 s/mm² (mean, 1233 ± 463 s/mm²), Figure 8. Qualitative assessment, where readers selected an optimal b-value to maximize lesion visibility on computed images for each patient case, showed the optimal b-value ranging from b = 567 to 2267 s/mm² (overall mean across all 3 readers, 1411 ± 383 s/mm²), (Figure 8). The ICC for agreement on optimal b-value between readers was 0.59 (95% CI: 0.33 to 0.78). The ICC for agreement between reader-selected optimal b-values (qualitative assessment) and optimal b-values based on peak CNR (quantitative assessment) was 0.50 (95% CI: 0.18 to 0.73). On average, readers preferred a b-value that was 178 s/mm² higher (95% CI: 26 to 331 s/mm², p=0.023) than the b- value exhibiting peak CNR.

Figure 7.

Two patients exhibiting differing variations of lesion CNR with b-value. Shown are a) a 36-year-old woman with a 20 mm invasive ductal carcinoma and heterogeneously dense fibroglandular tissue, where the mass-type lesion is more visible and distinct from normal tissue at higher b-values (peak CNR at b = 2500 s/mm² for acquired DWI and b = 1600 s/mm² for computed DWI) and b) a 56-year-old woman with a 31 mm invasive ductal carcinoma and scattered fibroglandular tissue, where the non-mass enhancement-type lesion is more visible at lower b-values (peak CNR at b = 0 s/mm²). Arrows indicate the malignancy on the reference DCE MRI, acquired and computed diffusion-weighted images.

Figure 8.

Individual optimal b-values across 30 patients by quantitative and qualitative assessment. Shown are a) distribution of b-values found to maximize lesion conspicuity as measured by CNR (quantitative assessment), b) distribution of b-values selected to maximize lesion conspicuity, averaged across three readers (qualitative assessment), and c) boxplots comparing optimal b-values based on CNR and reader assessment.

Association with Patient and Tumor Characteristics

Post-hoc comparison of optimal b-values, based on both qualitative and quantitative assessment, with clinical factors identified significant associations with both breast density and age (Table 2). The optimal b-value by qualitative assessment for patients with dense breasts (heterogeneous FGT or extremely dense; mean b = 1494 ± 333 s/mm²) was higher than that for patients with less dense breasts (scattered FGT or fatty; mean b = 1000 ± 377 s/mm², p=0.015). Additionally, the optimal b-value in women ≤ 50 years (mean b = 1538 ± 377 s/mm²) was higher than that in older women (mean b = 1158 ± 257 s/mm², p=0.002). Similar trends were observed for optimal b-values based on quantitative assessment of peak CNR but did not reach statistical significance. In terms of lesion characteristics, masses demonstrated higher mean optimal b- values than non-mass enhancement both by quantitative and qualitative assessment, although the differences did not reach significance (p = 0.12 and 0.13, respectively). No other significant associations were identified between optimal b-value and other clinical factors (i.e. lesion size, histology, and Nottingham grade).

Table 2.

Associations of optimal b-values with patient and tumor characteristics.

Variable	N	Quantitative Assessment (Maximal CNR)		Qualitative Assessment (Maximal Visibility)
		b-value*	P-value†	b-value*	P-value†
Age
≤ 50 years	20	1335 ± 479	0.12	1538 ± 377	0.002
> 50 years	10	1030 ± 371		1158 ± 257
Breast density
Dense	25	1296 ± 413	0.070	1494 ± 333	0.015
Not dense	5	920 ± 622		1000 ± 377
Size
≤ 2 cm	17	1341 ± 469	0.16	1424 ± 401	0.90
> 2 cm	13	1092 ± 433		1396 ± 374
Lesion type
Mass	25	1296 ± 448	0.12	1459 ± 375	0.13
NME	5	920 ± 455		1173 ± 364
Histology
IDC	26	1254 ± 485	0.56	1426 ± 406	0.38
ILC or DCIS	4	1100 ± 294		1317 ± 174
Nottingham grade
1–2	13	1377 ± 482	0.22	1363 ± 425	0.63
3	15	1127 ± 446		1429 ± 375

^*Values are mean ± SD of the per-patient optimal b-values (s/mm²)

^†Wilcoxon rank-sum test comparing optimal b-values between two groups

Abbreviations: NME = non-mass enhancement, IDC = invasive ductal carcinoma, ILC = invasive lobular carcinoma, DCIS = ductal carcinoma in situ

DISCUSSION

Our study demonstrates that conspicuity of breast malignancies on unenhanced diffusion- weighted MRI generally increases with increasing b-value and that peak lesion conspicuity may occur at b-values higher than those used in typical breast DWI protocols (e.g., 600 to 1000 s/mm² (22)). Furthermore, our data suggests optimal b-values for cancer detection on DWI tend to be higher in younger women and those with dense breasts. If DWI is to be used as an unenhanced screening tool, high b value acquisitions or ability to compute high b images for interpretation will be important to maximize sensitivity for cancer detection for these populations. Computed high b-value diffusion-weighted images allow improved lesion conspicuity without extending scan times and can provide flexible post-hoc selection of optimal diffusion-weighting.

Increasing lesion CNR with b-value was generally observed for both acquired and computed images. Averaged across all lesions, peak CNR occurred at b=1200 s/mm² for computed images and 1500 s/mm² for acquired images. This difference may relate to the limited sampling of b-values acquired during patient scans (b = 0, 100, 800, 1500, 2500 s/mm²), which did not include images acquired at b=1200 s/mm². Our findings for computed DWI agree well with those of Han et al. who showed in a previous study using DWI acquired over many b-values that the CNR of malignant lesions peaked at 1200 s/mm² (9).

In our study, CNR was slightly higher for acquired versus computed images at the same b-value. During DWI acquisition, high b-value averaging is commonly used to improve SNR. High b-value averaging increases the number of signal averages acquired for each image, increasing scan time accordingly. Synthesized or computed high b-value images are computed using images acquired at lower b-values with less signal averages. The differences in lesion CNR observed between computed and acquired high b images are likely due to in part to the absence of averaging occurring in computed images compared to acquired images, as demonstrated in the phantom experiments. Similar findings were made by Tamura et al. who reported that increasing the number of signal averages for acquired images used to compute high b-value images could increase the tumor conspicuity on computed images (15). Another important factor that may result in higher CNR in acquired images is related to non-Gaussian diffusion behavior at higher b-values. The diffusion of water in vivo is not accurately modeled with a monoexponential decay curve as it assumes free diffusion (and resulting Gaussian distribution) of water whereas in biologic tissues, diffusion is restricted due to the presence of barriers such as cell membranes.

Therefore, in some cases, ADC may not be able to accurately predict the behavior of water in tissue microstructure. Alternatively, the diffusion kurtosis model, first described by Jenson et al, can better reflect the microstructural complexity of tissue structures (23). This model has two parameters: a diffusion coefficient and a kurtosis term. Higher kurtosis values imply more impediments to normal, Gaussian diffusion and greater complexity within the imaged system (24), more commonly observed in malignant tissues vs. benign breast tissues (25). However, by computing images using a monexponential decay model, the biological kurtosis effect that can cause signal to be further retained in breast malignancies at higher b-values (>1500 s/mm²) is ignored.

Qualitative assessment by three readers found no significant difference in lesion visibility for acquired vs. computed b=1500 s/mm² images. These findings differ somewhat from those of Park et al who reported higher reader sensitivity using computed vs. acquired b=1500 s/mm² images for detecting breast cancers in a blinded reader study, which they attributed to higher SNR of computed images (14). These differences in study findings likely relate to substantially higher SNR for acquired b = 1500 s/mm² images in our study due to differences in DW MRI imaging protocols, as our study utilized a higher number of signal averages (NSA = 6 vs. NSA = 1 for the Park et al study), shorter TE (79.9 ms vs. 130 ms), and greater voxel size (1.8×1.8×4 mm = 12.96 mm³ vs. 1.5×1.5×3 mm = 6.75mm³). Park et al also observed a higher false positive rate for computed images, reportedly due to artifacts resulting from patient motion or distortion effects causing spatial inaccuracies (misregistration) between the two low b-value images used to synthesize the high b images. Anecdotally, such detrimental artifacts were not observed in our unblinded study, although we did not specifically investigate incidental findings and therefore this potential increase in false positives on computed high b diffusion images warrants further investigation.

The optimal b-value for lesion conspicuity varied case-to case by both quantitative and qualitative assessment. The association of optimal b-value with both age and breast density supports similar findings by Bickel et al (7) and suggests that use of higher b-values can improve cancer detection on unenhanced DWI for younger women with dense breasts. Our findings further suggest that masses may be best visualized at higher b-values, while non-masses, may be less apparent at very high b-values, which should also be considered in developing future DWI interpretation strategies. Although other tumor characteristics such as lesion size, histology, and grade did not significantly influence optimal b-values in our study, larger patient cohorts are needed to better evaluate these associations. Ultimately, a variety of factors contribute to lesion conspicuity, as illustrated by the fact that radiologists tended to prefer higher b-values than those quantitatively determined to produce peak CNR. This may relate to characteristics such as breast composition, lesion and parenchymal heterogeneity, and image quality factors (e.g., SNR, presence of artifacts, and quality of fat suppression) that can also affect overall lesion visibility.

There are limitations to our study. Our accuracy for identifying the b-value corresponding to peak CNR for acquired images was limited due to the select number of higher b-values that were acquired (due to scan time considerations). Though we observed a CNR peak for acquired images at 1500 s/mm², we expect that if more b-values were acquired, the CNR would peak between 800–1500 s/mm², similarly to the observed peak for computed images. Since our study was limited to 30 patients, further investigation in a larger clinical cohort is needed to confirm our findings and to better understand patient and clinical factors influencing optimal b values and cancer conspicuity. Furthermore, all patients had known diagnosis of malignancy at the time of MRI. Post biopsy changes could influence appearance of lesions. Additionally, radiologists were not blinded to the location or presence of lesions in order to select optimal b-value and judge lesion conspicuity. As a result, this study did not evaluate the sensitivity and specificity of computed high b-value images for detecting breast cancer.

In conclusion, our study suggests overall conspicuity of breast malignancies on DWI is maximized using b-values ranging 1200 s/mm² to 1400 s/mm², which are higher than those used in typical multiparametric breast MRI protocols. A recent consensus by an international breast DWI working group has suggested using a high b value of 800 s/mm² for standardized breast DWI acquisitions and for robust lesion ADC measures (26). However, results of our study and others suggest that if the intended application of DWI is for unenhanced screening, even higher b-values of ≥1200 s/mm² should also be acquired or computed (7). Computing high b-value images rather than acquiring them directly could produce images at multiple diffusion- weightings and improve lesion conspicuity while saving scan time. Although computed DWI may incur a slight reduction in CNR versus directly acquired images, there was no significant impact on lesion conspicuity based on visual assessment. Our study further suggests that optimal b-value for lesion conspicuity differs with breast density and other patient characteristics, and therefore computed high b-value DWI offers the added benefit of allowing readers the flexibility to adjust b-value at the time of interpretation.

Abbreviations

MRI

magnetic resonance imaging

DCE

dynamic contrast-enhanced

DWI

diffusion-weighted imaging

CNR

contrast-to-noise ratio

FGT

fibroglandular tissue

SNR

signal-to-noise ratio

NSA

number of signal averages

ADC

apparent diffusion coefficient

ICC

intraclass correlation coefficient

ROI

regions of interest

CI

confidence interval