Institutional Variability in Ultrafast Breast MR Imaging: Comparing Compressed Sensing and View Sharing Techniques with Different Patient Populations and Contrast Injection Protocols

Abstract

Purpose

To assess the institutional variability in ultrafast dynamic contrast-enhanced (UF-DCE) breast MRI using time-resolved angiography with stochastic trajectories (TWIST)-volumetric interpolated breath-hold examination (VIBE) and compressed sensing (CS)-VIBE sequences acquired at 2 different institutions with different patient populations and contrast injection protocols.

Methods

UF-DCE MR images of 18 patients from site A acquired using a TWIST-VIBE sequence, and UF-DCE MR images of 18 patients from site B acquired with a CS-VIBE sequence, were retrospectively evaluated and compared. The 2-site patient cohort was matched for patient age, background parenchymal enhancement, malignancy or benignity, and lesion size. Qualitative assessments included noise, blurring, poor fat suppression, aliasing artifact, motion artifact, lesion conspicuity, lesion morphology, time-intensity-curve smoothness, and vessel delineation. For quantitative assessment, the bolus arrival time was evaluated for each lesion, and its diagnostic performance in discriminating between benign and malignant lesions was examined using receiver operating characteristics analysis.

Results

Thirteen malignant and five benign lesions were included from each site. Qualitative evaluation revealed that poor fat suppression and aliasing artifacts were visible in images from site A with TWIST-VIBE (P = 0.004 and P < 0.001), whereas motion artifacts were present in images from site B with CS-VIBE (P = 0.04). Lesion morphology assessments (P < 0.001) and vessel delineation (P < 0.001) were superior for images from site B with CS-VIBE. Bolus arrival time was significantly longer with TWIST-VIBE than with CS-VIBE, for both benign and malignant lesions (P < 0.001). The area under the receiver operating characteristics curve was 0.55 for site A and 0.69 for site B (P = 0.39).

Conclusion

Both acquisitions allowed evaluation of breast lesions with good lesion conspicuity and time-intensity-curve smoothness, whereas CS-VIBE was superior to TWIST-VIBE for morphological evaluation of breast lesions and depiction of blood vessels in the breast. Injection rate appears to have a significant impact on semi-quantitative parameters derived from UF-DCE MRI.

Introduction

The utility of ultrafast dynamic contrast-enhanced (UF-DCE) MRI for evaluation of breast lesions has been previously demonstrated.^1–8 UF-DCE MRI typically refers to high-speed imaging with a temporal resolution of less than approximately 7s, with multiple phases of contrast enhancement being acquired in the first minute after contrast injection to capture the kinetic information of contrast influx. UF-DCE MRI-derived parameters, such as maximum slope (MS) and time-to-enhancement (TTE), can be used to discriminate malignant from benign breast lesions with similar diagnostic performance to conventional time-intensity curve (TIC) analysis using early and delayed phase imaging.^1–4 Another benefit of using UF-DCE MRI is improved detection of breast lesions when background parenchymal enhancement (BPE) is pronounced.^5,6 In the context of neoadjuvant treatment, UF-DCE MRI was reported to be better than conventional DCE MRI at predicting the pathological complete response of breast cancer after neoadjuvant systemic therapy.⁷

Acceleration techniques such as parallel imaging, view-sharing, and compressed sensing (CS) are used to improve temporal resolution when acquiring breast UF-DCE MRI. Parallel imaging has limited acceleration capability because of limits on the number of coils, and therefore the latter 2 methods have recently been widely used.^4,8 View-sharing is a technique that accelerates image acquisition by skipping some of the data that are normally acquired and recovering the missing data points from surrounding data points at different k-space locations and/or time points (view-shared). Several sequences, such as k-space-weighted image contrast (KWIC) and time-resolved angiography with stochastic trajectories (TWIST), are widely used in UF-DCE MRI of the breast. In contrast, CS accelerates acquisition by reconstructing images from highly under-sampled data, assuming that the data are sparse in a certain transform design. It has been applied to medical imaging since 2007,⁹ and high contrast images, including contrast-enhanced breast MRI, are considered a good application.¹⁰ CS is reported to provide images with better spatial resolution than view-sharing, but needs longer reconstruction times.¹¹

There is currently no standard protocol for UF-DCE MRI, and several imaging parameters, including acceleration methods, use of fat-suppression (or subtraction), and speed of contrast-agent injection, vary across institutions. Semi-quantitative parameter values also vary widely across studies, which may be due to these differences in acquisition.⁸ Understanding the variability in UF-DCE MRI sequences and its impact on lesion evaluation is essential for interpreting results and is an important step toward integrating the evidence on the clinical value of UF-DCE MRI. Therefore, the aim of this study was to compare UF-DCE MRI between 2 institutions with different imaging methods.

Materials and Methods

This retrospective study was approved by the institutional review board of Kyoto University Hospital (R0610-7) and the requirement for obtaining written informed consent was waived.

Study population

A total of 41 patients underwent contrast-enhanced MRI including a TWIST-volumetric interpolated breath-hold examination (VIBE) at site A between August and December 2023. We excluded patients without a pathological diagnosis, without a lesion large enough to place a 3 × 3 mm ROI, patients undergoing chemotherapy, and recurrent cases, leaving 18 patients for evaluation. During the same period, 155 patients underwent contrast-enhanced MRI including CS-VIBE and pathological diagnosis at site B. For the comparison with patients from institution A, 18 patients from site B were selected by patient matching. Using the “MatchIt” package in R (version 4.2.1, The R Foundation for Statistical Computing, Vienna, Austria), nearest 1:1 matching was enforced for the following categorical or numerical covariates obtained from electronic records: age, BPE (minimal, mild, moderate, or marked), malignancy or benignity, and lesion size on contrast-enhanced MRI. A flowchart detailing the patient inclusion and exclusion is provided in Fig. 1.

Fig. 1.

Flowchart of the patient inclusion and exclusion criteria. BPE, background parenchymal enhancement; CS, compressed sensing; TWIST, time-resolved angiography with stochastic trajectories; VIBE, volumetric interpolated breath-hold examination.

Image acquisition

At site A, all MRI examinations were performed using a 3T unit (MAGNETOM Skyra; Siemens Healthineers, Erlangen, Germany) and 18-channel dedicated breast coils. Patients were placed in a prone position and both upper extremities were placed alongside the body. Gadobutrol (Gadovist; Bayer AG, Leverkusen, Germany) (0.1 mL/kg) was intravenously injected at the rate of 1 mL/s, followed by 50 mL of saline at the same rate. UF-DCE MRI was obtained 15s before to 60s after contrast injection using a TWIST-VIBE sequence (TR 6.4 ms; TE 3.08 ms; flip angle 10 degree; size of fully sampled central k-space region A, 30%; sampling density of peripheral k-space region B, 33%¹²; forward reconstruction; time resolution 8.5s; time to center 3.04 + 8.46x; voxel size 0.7 × 0.7 × 2.5 mm; 8 phases; water-excitation for fat suppression) (Table 1). Other DCE-MRI sequences included pre-contrast, initial phase (60–120s after contrast injection), high spatial-resolution contrast-enhanced MRI (120–300s after contrast injection), and delayed phase (300–360s after contrast injection) (Supplementary Table).

Table 1.

Image acquisition

	Site A	Site B
Scanner	MAGNETOM Skyra (3T)	MAGNETOM Prisma (3T)
Coil	18ch breast coil	18ch breast coil
Position	Prone with upper extremities alongside the body	Prone with upper extremities elevated
Contrast agent	Gadobutrol	Gadobutrol
Injection speed, ml/s	1	2
Sequence	TWIST-VIBE	CS-VIBE
TR, ms	6.4	4.8
TE, ms	3.08	2.46
FA, degree	10	15
Time resolution, s	8.5	3.7
Time to center, s	3.04 + 8.46x	2.07 + 3.65 (x-1/2)
Voxel size, mm	0.7 × 0.7 × 2.5	0.9 × 0.9 × 2.5
Fat suppression	water-excitation	none

CS, compressed sensing; TWIST, time-resolved angiography with stochastic trajectories; VIBE, volumetric interpolated breath-hold examination.

At site B, all MRI examinations were also performed using a 3T unit (MAGNETOM Prisma, Siemens Healthineers) and 18-channel dedicated breast coils. Patients were placed in a prone position and both upper extremities were elevated. Gadobutrol (0.1 mL/kg) was intravenously injected at the rate of 2 mL/s, followed by 20 mL of saline at the same rate. UF-DCE MRI was obtained 15s before to 60s after contrast injection using a prototype CS-VIBE sequence (TR 4.8 ms; TE 2.46 ms; flip angle 15 degree; time resolution 3.7s; time to center 2.07 + 3.65 (x − 1/2)s; voxel size 0.9 × 0.9 × 2.5 mm; 20 phases; no fat suppression) (Table 1). CS reconstruction took approximately 10 mins after the completion of the acquisition and was usually conducted after the completion of the whole dynamic acquisition. Other DCE-MRI sequences included pre-contrast, initial phase (60–120s after contrast injection), high spatial-resolution contrast-enhanced MRI (120–300s after contrast injection), and delayed phase (300–360s after contrast injection) (Supplementary Table).

Image evaluation

Image evaluation was performed by 2 breast radiologists (M.H. and M.K. with 10 years and 25 years of experience, respectively) 3 months after patient selection. The radiologists were blinded to the final diagnosis, although the site of the target lesions was notified to the readers to ensure that they evaluated the same lesions as those that were pathologically diagnosed. For CS-VIBE at site B, subtracted images were used for the evaluation. Disagreements in the evaluations between the 2 readers were resolved by consensus.

Qualitative evaluation: Using a workstation (Aquarius NET Viewer; TeraRecon, Foster City, CA, USA), 9 items (noise, blurring, poor fat suppression, aliasing artifact, motion artifact, lesion conspicuity, lesion morphology, TIC smoothness, and vessel delineation) were evaluated in consensus for each MRI, using a 5-point scale (1: poor to 5: excellent). For lesion morphology assessment, we investigated the feasibility of Breast Imaging Reporting & Data System (BI-RADS) morphology evaluation using images from the final phase of ultrafast MRI. A score of 5 was assigned to cases where the morphology could be reliably evaluated as equivalent by referencing the initial phase of DCE images. Lesion conspicuity was assessed using both the final phase images and the maximum intensity projection (MIP) of the final phase subtracted images. Vessel delineation was evaluated on MIP of the final phase subtracted images. Other items were evaluated on the final phase images.

Quantitative evaluation: The same 2 radiologists also evaluated the bolus arrival time (BAT) of each lesion using the same workstation.^13,14 The BAT was determined as the time-to-center of the phase at which lesions started to enhance. The beginning of the enhancement was judged visually using MIPs and original images, or if visual judgment was difficult, it was determined as the phase when the lesion signal exceeded 20% of the peak signal on TICs.

Statistical analysis

Clinical factors were compared between sites A and B using Fisher’s exact test, Pearson’s Chi-squared test, and the Wilcoxon rank-sum test. All categorical variables were compared using the Brunner–Munzel test, and continuous variables were compared using Welch’s t test. For all tests, a P < 0.05 was considered statistically significant. Receiver operating characteristics (ROC) analysis was used to examine the diagnostic performance of BAT in the discrimination between benign and malignant lesions. Bootstrap operations, performed using the pROC package, were used to calculate confidence intervals (CIs) of ROCs curves and to compare them.¹⁵ All statistical analyses were performed using open-source software (R-4.2.1; R Foundation for Statistical Computing).

Results

For each institution, 5 benign and 13 malignant lesions were included. The histological diagnoses of the benign lesions were 3 intraductal papilloma, 1 usual ductal hyperplasia, and 1 adenosis at site A, and 1 intraductal papilloma, 1 adenosis, 1 lymphocytic mastitis, 1 benign phyllodes, and 1 usual ductal hyperplasia at site B. Malignant lesions included 8 invasive ductal carcinoma (all luminal subtype), 3 invasive lobular carcinoma (all luminal subtype) and 2 ductal carcinoma in situ at site A, and 10 invasive ductal carcinoma (9 luminal and 1 triple negative subtype), 2 ductal carcinoma in situ and 1 invasive lobular carcinoma (luminal subtype) at site B. Table 2 shows the patient characteristics.

Table 2.

Patient characteristics at each site

Characteristics	Overall, N = 361	Site A, N = 181	Site B, N = 181	P-value2
Age	57 (11)	57 (11)	57 (12)	0.9
BPE				>0.9
Minimal	14/36 (39%)	7/18 (39%)	7/18 (39%)
Mild	9/36 (25%)	4/18 (22%)	5/18 (28%)
Moderate	9/36 (25%)	5/18 (28%)	4/18 (22%)
Marked	4/36 (11%)	2/18 (11%)	2/18 (11%)
Size (mm)	21 (17)	21 (19)	22 (14)	0.4
Benign/Malignant				>0.9
Benign	10/36 (28%)	5/18 (28%)	5/18 (28%)
Malignant	26/36 (72%)	13/18 (72%)	13/18 (72%)
Mass/NME				>0.9
Mass	30/36 (83%)	15/18 (83%)	15/18 (83%)
NME	6/36 (17%)	3/18 (17%)	3/18 (17%)

¹Mean (SD); n/N (%).

²Wilcoxon rank sum test; Fisher’s exact test; Pearson’s Chi-squared test.BPE, background parenchymal enhancement; NME, non-mass enhancement.

Qualitative evaluation revealed that poor fat suppression and aliasing artifacts could be seen in images from site A (P = 0.004 and P < 0.001) and that motion artifacts could be seen in images from site B (P = 0.04). Lesion morphology assessment (P < 0.001) and vessel delineation (P < 0.001) were superior for images from site B with CS-VIBE. No significant differences were observed for the other items.

In the quantitative evaluation, 1 patient from site B was excluded because the lesion was not enhanced on UF-DCE MRI. BAT was significantly longer in the TWIST-VIBE images than in the CS-VIBE images, for both benign and malignant lesions (P < 0.001) (Table 3). The area under the ROC curve was 0.55 (95%CI: 0.28–0.77) for site A and 0.69 (95%CI: 0.43–0.90) for site B (P = 0.39). Representative images of breast cancer acquired using TWIST-VIBE and CS-VIBE are shown in Figs. 2 and 3, respectively. Both MIPs of the last phase (Fig. 2a and Fig. 3a) show mass lesions, whereas breast vessels are clearly visible in Fig. 3 acquired using CS-VIBE. When comparing the last phase images (Fig. 2b and Fig. 3b), the margin of the mass lesion is more clearly delineated in Fig. 3, acquired using CS-VIBE. The BAT in Fig. 2 acquiring with TWIST-VIBE was 53.8s, more than twice the BAT in Fig. 3 (21.9s), acquired with CS-VIBE.

Table 3.

Qualitative and quantitative values of site A and site B

Variables	Site A, N = 181	Site B, N = 181	P-value2
Noise	5.0 (0.0)	5.0 (0.0)	NaN
Blur	4.9 (0.2)	5.0 (0.0)	0.331
Insufficient fat suppression	4.5 (0.7)	5.0 (0.0)	0.004
Aliasing artifacts	4.1 (0.6)	5.0 (0.0)	<0.001
Motion artifacts	5.0 (0.0)	4.8 (0.4)	0.042
Lesion conspicuity	4.6 (0.9)	4.7 (1.2)	0.251
Morphological assessment	4.1 (0.7)	4.7 (1.2)	<0.001
TIC smoothness	4.6 (1.0)	4.4 (1.2)	0.215
Vessel delineation	2.1 (1.6)	4.2 (1.5)	<0.001
BAT
Malignant (s)	49.9 (5.6)	24.5 (4.8)	<0.001
Benign (s)	50.4 (4.7)	27.4 (2.1)	<0.001
AUC (95%CI)	0.55 (0.28-0.77)	0.69 (0.43-0.90)	0.39

¹Mean (SD).

²Brunner-Munzel test; Welch’s t test; Boostrap operations.

AUC, area under the receiver operating characteristics curve; BAT, bolus arrival time; CI: confidence interval; TIC, time-intensity curve.

Fig. 2.

Representative time-resolved angiography with stochastic trajectories-volumetric interpolated breath-hold examination images of left invasive breast cancer (luminal subtype, histological grade 1, Ki-67 labeling index: 15%). Maximum intensity projection of the last phase subtracted images (time to center, 70.7s) (a) shows a mass lesion in the left breast, whereas breast vessels are indistinct. In the final phase image (time to center, 70.7s) (b), the mass lesion is clearly delineated, but the margin is somewhat unclear compared with the findings in the initial phase (c). A smooth time-intensity-curve could be drawn using a 3 × 3 mm region of interest and the bolus arrival time was 53.8s (d).

Fig. 3.

Representative compressed sensing-volumetric interpolated breath-hold examination images of right breast cancer with apocrine differentiation (triple negative subtype, histological grade 2, Ki-67 labeling index: 5%). Maximum intensity projection of the 18th phase subtracted images (time to center, 65.9s) (a) shows a mass lesion in the right breast with developed breast vessels. In the final phase subtracted image (time to center, 73.2s) (b), the mass lesion is clearly delineated and the margin and internal signal can be evaluated, as they could in the initial phase (c). A smooth time-intensity-curve could be drawn using a 3 × 3 mm ROI and the bolus arrival time was 21.9s (d).

Discussion

In this study, matched cohorts were used to compare the image quality of UF-DCE breast MRI between TWIST-VIBE and CS-VIBE sequences acquired at 2 different institutions. CS may be superior to view-sharing for morphological evaluation of breast lesions and depiction of blood vessels in the breast, whereas lesion conspicuity and TIC smoothness were comparable between the 2 acquisitions.

A previous study reported that in UF-DCE MRI of the breast, higher spatial resolution was obtained with CS than with view-sharing.¹¹ In the current study, higher in-plane resolution was adopted at site A with TWIST-VIBE, at the cost of lower temporal resolution, but the morphological evaluation of breast lesions was still inferior on images from site A than on CS-VIBE images from site B. Temporal blurring in the view sharing method may lead to a reduction in morphological clarity. In the present study, the slower injection rate at site A might also be responsible for reduced clarity of enhancement. In contrast, CS-VIBE first reconstructs images within each time phase, and there is no mixing of signals from different time phases at this point. Images are then processed in the time direction to remove artifacts and noise, but the effect of this processing on image quality is expected to be less than it is with TWIST-VIBE.

We found that breast vessels were clearly delineated with CS-VIBE at site B, as in previous studies,^16,17 and this delineation was significantly better than that on TWIST-VIBE images acquired at site A. As mentioned above, the higher spatial resolution and injection rate may have led to the better conspicuity of thin but well-enhanced breast vessels with CS-VIBE. Although evaluation of breast vessels is not mandatory at this time, some previous studies have proposed utilizing breast vessels to differentiate breast lesions. The time interval between arterial and venous visualization^16,17 and the number of tumor-associated vessels on MRI¹⁸ have been shown to discriminate malignant from benign breast lesions. The vasculature of cancerous breasts on MIPs was also shown to correlate with the response of breast cancer to neoadjuvant chemotherapy.¹⁹ Therefore, the ability to evaluate breast vessels may be an advantage of CS over view-sharing, although differences in injection speed may also have affected the current results.

Ghosting artifacts and infolding artifacts are reported to be more common in TWIST acquisitions, whereas a specific pulsation artifact has previously been observed in CS,¹¹ but this trend was not seen in the current study. Differences in artifacts in the current results may be due to the position of the patient at the time of imaging and the presence or absence of fat suppression. The higher prevalence of aliasing artifacts in images from site A may be due to the placement of the upper extremities along the body. Subtraction of CS-VIBE images without fat suppression may result in a higher prevalence of motion artifacts in images from site B, but not inadequate fat suppression.

There was no significant difference in the smoothness of TICs between the 2 institutions. View sharing can affect the TIC because it shares data from different time points, but TWIST uses data from adjacent time points, and therefore the TIC may have remained smooth. In contrast, CS-VIBE first performs image reconstruction within each time phase, so there is no mixing of signals at this point.

BAT was significantly longer in images from site A than in images from site B, which may be attributed to the injection rate and temporal resolution. Other semi-quantitative parameters might also be affected by the injection rate, the contrast agent used and temporal resolution, although they have not been considered in this study. Therefore, both the injection rate, the contrast agent used and temporal resolution should be considered when interpreting TICs from UF-DCE MRI. The diagnostic performance in the discrimination between benign and malignant lesions was better at site B than at site A, although there were no significant differences. This may be attributed to the time resolution and injection rate, albeit the 2 sites imaged different patient samples. A recent review article showed that a time resolution shorter than 5s was associated with greater specificity when differentiating benign and malignant breast lesions using MS and TTE.³ In the current study, BAT, which is considered easier to measure and more versatile than MS and TTE, was introduced for the quantitative evaluation. The present results suggest that even with BAT, a temporal resolution shorter than 5s may improve the ability to distinguish benign from malignant breast lesions.

This study has several limitations. First, the study cohort was small and was not the same between the 2 sites. However, we used exact matching to eliminate differences in patient and lesion characteristics as much as possible. Second, there were many differences between the imaging methods of the 2 facilities, making it difficult to interpret the results. Further study with better-aligned imaging conditions, including injection rate, should clearly reveal the differences between view-sharing and CS.

Conclusion

In conclusion, both CS and view-sharing allowed evaluation of breast lesions with good lesion conspicuity and TIC smoothness, whereas CS was superior to view-sharing for morphological evaluation of breast lesions and depiction of blood vessels in the breast, which may lead to more accurate diagnosis and classification of breast lesions with UF-DCE MRI. Additionally, injection rate appears to have a significant impact on semi-quantitative parameters extracted from UF-DCE MRI.

Supplementary Materials

Supplementary Table.

The MRI sequence parameters