Ultrafast dynamic contrast-enhanced breast MRI may generate prognostic imaging markers of breast cancer

The institutional review board approved our retrospective study and waived the need for written informed consent. We conducted this study in compliance with the Health Insurance Portability and Accountability Act.

From February 2017 to January 2018, contrast-enhanced breast 3T MRI examinations routinely included ultrafast DCE-MRI along with standard DCE-MRI as part of a hybrid DCE-MRI protocol in our institution; however, clinical assessments including BI-RADS categorization were made using standard DCE-MRI without ultrafast DCE-MRI. We conducted a retrospective search of the institutional electronic medical record during this period to identify examinations that met the following inclusion criteria: (1) performed with a hybrid DCE-MRI protocol on a single 3T scanner with a 16-channel breast coil, (2) depicted pathologically proven breast lesions, (3) categorized as BI-RADS MRI 4–6, and (4) not performed for post-treatment evaluation after chemotherapy or surgery. Of 2,351 consecutive contrast-enhanced breast 3T MRI examinations including screening and diagnostic examinations, 196 examinations met the inclusion criteria. We included all pathologically proven breast lesions depicted on these examinations with the exception of the following lesions: lesions pathologically diagnosed as a special type malignancy (n = 2, malignant phyllodes tumor, 1; spindle cell sarcoma, 1), lesions pathologically diagnosed as benign lesions (n = 79), lesions without a one-to-one pathological diagnosis (n = 36), lesions with no or minimal residual enhancement difficult to differentiate from post-biopsy change (n = 19; invasive carcinoma, 15; ductal carcinoma in situ [DCIS], 4), and lesions with severe patient motion during MRI scanning which could not be resolved by motion correction technique (n = 7; invasive carcinoma, 6; DCIS, 1). In total, 125 examinations with 142 pathologically proven breast cancer lesions were included in this study (Fig. 1). The current study included a partial overlap in the study cohort, with separate studies investigating the diagnostic performance of ultrafast DCE-MRI-derived parameters [15] and the efficacy of radiomic analysis using standard DCE-MRI for sub-1 cm lesions [16]: 97 examinations with 106 lesions and 74 examinations with 79 lesions, respectively. Detailed information of the overlap is shown in the Supplemental Table.

All patients underwent MRI examinations in the prone position on a single scanner: 3.0T MRI system (Discovery 750, GE Medical Systems, Waukesha, WI) with a dedicated 16-channel breast coil (Sentinelle, Invivo, Gainesville, FL). We used the same MRI protocol as in a previous study [15]. Briefly, one pre-contrast phase and three post-contrast phases of standard DCE-MRI, and 15 phases of ultrafast DCE-MRI were acquired. After pre-contrast imaging was acquired for standard DCE-MRI, ultrafast DCE-MRI was acquired continuously for 15 timepoints during the first approximately 60 s, starting simultaneously with the start of contrast injection. The contrast agent (Gadavist; Bayer Healthcare Pharmaceuticals Inc., Whippany, NJ) was administered at a concentration of 0.1 mmol gadobutrol per kg body weight and a rate of 2 ml/s, followed by a 40 ml saline flush at the same rate. Immediately after ultrafast DCE-MRI, standard DCE-MRI was acquired continuously at three timepoints (Fig. 2). The acquisition parameters were as follows: for ultrafast DCE-MRI using a 3D fat-suppressed T1-weighted differential sub-sampling with cartesian ordering (DISCO) sequence, TR = 3.8 ms, TE1 and TE2 = 1.1/2.2 ms, flip angle = 12°, field of view = 34 cm × 34 cm, acquired matrix = 212 × 212, in-plane spatial resolution = 1.6 × 1.6 mm, thickness = 1.6 mm, number of slices = 166 (or higher to ensure full breast coverage), bandwidth = ± 142.86 kHz, ARC acceleration factor = 4 (phase) × 2 (slice), temporal resolution = 2.7–4.6 s/phase, temporal foot-print (time to acquire a complete k-space for a phase) = 9 s, axial orientation. For standard DCE-MRI using a 3D fat-suppressed T1-weighted volume imaging breast assessment (VIBRANT) sequence, TR/TE = 7.9/4.3, flip angle = 12°, field of view = 34 cm × 34 cm, acquired matrix = 300 × 300, in-plane spatial resolution = 1.1 × 1.1 mm, adiabatic fat suppression, number of slices = 190 (depending on coverage), bandwidth = ± 62.5 kHz, ASSET acceleration factor = 2 (phase) × 1 (slice), scan time = ~ 120 s/phase, axial orientation.

Radiologists 1 and 2 (NO and ESK with 7 and 13 years of experience in breast MRI, respectively) reviewed the first post-contrast standard DCE-MR images for all lesions and evaluated the BI-RADS MRI lesion type (mass, non-mass enhancement [NME], and focus) in consensus. The radiologists were blinded to pathological reports.

Histopathology was retrospectively determined; a careful review of pathological reports of both core biopsy specimens and surgical specimens was performed to acquire one-to-one pathological diagnoses of the lesions depicted on MRI. All lesions were classified into invasive carcinoma or DCIS. Microinvasive carcinomas were classified as DCIS because they consist primarily of DCIS. Invasive carcinomas were further classified based on their tumor grade (grade 1, 2, or 3), their histopathological subtype (invasive ductal carcinoma [IDC], invasive lobular carcinoma [ILC], or mixed invasive ductal/lobular carcinoma [mixed IDC/ILC]), and their molecular subtype (triple negative type [hormone receptor negative, HER2 negative], HER2 type [HER2 positive regardless of the hormone receptor positivity/negativity], or luminal type [hormone receptor positive, HER2 negative]). Positive lymph node metastasis was pathologically determined by reviewing reports of either fine needle biopsy, sentinel lymph node biopsy, or axillary dissection. Oncotype DX recurrence score (Genomic Health, Redwood City, CA) was available if the lesion was diagnosed as luminal type, T1–2, and N0–N1mi (micrometastases) on surgical pathology. The recurrence risk was evaluated based on the score: low recurrence risk (score ≤ 17), intermediate recurrence risk (18 ≤ score ≤ 30), and high recurrence risk (score ≥ 31).

To assess the inter-reader reliability of MS and BAT, intraclass correlation coefficient was assessed using a cohort of 24 lesions randomly derived from the whole cohort, of which radiologist 3 (MCH, 10 years of experience in breast MRI) independently performed segmentation and calculated each parameter using the same method as radiologist 1. The intraclass correlation coefficient between the two radiologists was evaluated following the interpretation proposed by Cicchetti [20]: a value of 0.75–1.00 was considered excellent agreement; 0.60–0.74, good agreement; 0.40–0.59, fair agreement; and less than 0.40, poor agreement.

To test the hypothesis that MS and BAT may differ between invasive carcinoma and DCIS, we used generalized estimating equations (GEE) in R, accounting for the presence of multiple lesions per patient. The maximum cluster size present was three. The cluster size was effectively the number of lesions per patient. In invasive carcinomas, to test the hypothesis that MS and BAT may differ in invasive carcinomas according to cancer characteristics (tumor grade, histopathology, molecular subtype, lymph node metastasis status, and recurrence risk), we divided the invasive carcinomas into the respective two subgroups and compared MS and BAT between groups using GEE. For the cancer characteristics in which both MS and BAT were significantly different, we also performed exploratory multivariate logistic regression modeling using GEE. Area under the ROC curve (AUC) of the models was compared using the DeLong’s test [21].

In total, 142 pathologically proven breast cancer lesions in 125 female patients (mean age, 49.6 years; SD, 11.9 years; range, 21–79 years) were analyzed. Of the 125 patients, 111 had a single lesion, 11 had double lesions, and 3 had triple lesions. Four patients had bilateral cancers. Table 1 summarizes the detailed information of the patients including age, menopausal status, BI-RADS category, past history of breast cancer, and family history of breast or ovarian cancer. The median interval between breast MR imaging and core biopsy was 14.6 days (interquartile range of 8.6–21.8 days).

The 142 lesions comprised 96 mass (68%), 44 NME (31%), and 2 focus (1%) lesions. They were pathologically diagnosed as 124 invasive carcinomas (87%) and 18 DCIS (13%). The median size of the invasive carcinomas was 19.5 mm (interquartile range of 13–32.3 mm), and the median size of the DCIS was 13.5 mm (interquartile range of 9.3–19.5 mm). Table 2 summarizes the detailed information of the cancer characteristics including lesion type, lesion size, tumor grade, histopathology, molecular subtype, lymph node metastasis status, and Oncotype DX recurrence score.

The inter-reader reliability was excellent for MS with an inter-reader intraclass correlation coefficient of 0.951 (95% CI 0.891, 0.978) and excellent for BAT with an inter-reader intraclass correlation coefficient of 0.948 (95% CI 0.885, 0.977).

Invasive carcinomas showed significantly larger MS and shorter BAT than DCIS (P < 0.001 and P = 0.008, respectively). Table 3 summarizes these results and Fig. 3 shows the box and whisker plots.

Tumor grade 3 presented significantly shorter BAT than tumor grades 1–2 (P = 0.025) (Fig. 4). IDC presented significantly shorter BAT than ILC (P = 0.002) (Fig. 5). Triple negative or HER2 type presented significantly shorter BAT than luminal type (P < 0.001) (Fig. 6). Figure 7 shows images of representative cases. Neither of the parameters were significantly different between lymph node metastasis positive lesions and lymph node metastasis negative lesions or between lesions with a high/intermediate recurrence risk and lesions with a low recurrence risk. Table 4 summarizes these results.

We performed an exploratory multivariate logistic regression modeling for differentiating invasive carcinoma from DCIS in NME lesions (Table 5). We selected MS, BAT, and size as variables. Of the three variables, only BAT was significantly predictive of invasive carcinoma (P = 0.022). The AUC of the MS + BAT + size model was significantly higher than that of the size alone model: 0.80 vs. 0.56, P = 0.024 (Fig. 8).