Radiomics for Tumor Characterization in Breast Cancer Patients: A Feasibility Study Comparing Contrast-Enhanced Mammography and Magnetic Resonance Imaging

The aim of our intra-individual comparison study was to investigate and compare the potential of radiomics analysis of contrast-enhanced mammography (CEM) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) of the breast for the non-invasive assessment of tumor invasiveness, hormone receptor status, and tumor grade in patients with primary breast cancer. This retrospective study included 48 female patients with 49 biopsy-proven breast cancers who underwent pretreatment breast CEM and MRI. Radiomics analysis was performed by using MaZda software. Radiomics parameters were correlated with tumor histology (invasive vs. non-invasive), hormonal status (HR+ vs. HR−), and grading (low grade G1 + G2 vs. high grade G3). CEM radiomics analysis yielded classification accuracies of up to 92% for invasive vs. non-invasive breast cancers, 95.6% for HR+ vs. HR− breast cancers, and 77.8% for G1 + G2 vs. G3 invasive cancers. MRI radiomics analysis yielded classification accuracies of up to 90% for invasive vs. non-invasive breast cancers, 82.6% for HR+ vs. HR− breast cancers, and 77.8% for G1+G2 vs. G3 cancers. Preliminary results indicate a potential of both radiomics analysis of DCE-MRI and CEM for non-invasive assessment of tumor-invasiveness, hormone receptor status, and tumor grade. CEM may serve as an alternative to MRI if MRI is not available or contraindicated.


Introduction
Breast cancer is the most common cancer in women, accounting for 30% of all new cancer diagnoses in women [1]. Breast cancer prognosis is related to different factors, such as histopathologic subtypes, immunohistochemical (IHC) receptor status, and tumor grade [2][3][4], which are typically assessed with invasive tissue sampling. However, biopsies provide mere snapshots of tumor biology and are

Materials and Methods
In this HIPAA-compliant retrospective institutional review-board-approved( Code 12-214, Date 10/2012, Institutional Review Board Memorial Sloan Kettering Cancer Center) single-institution study, informed consent was waived.

Study Cohort
We included 118 patients over 21 years of age who underwent CEM and breast MRI between January 2010 and April 2018 as part of their inclusion in two prospective CEM studies. Sixty patients with breast cancer were part of a prospective study [18,26] that investigated the utility of CEM for breast cancer detection and staging. The remainder of the patients were participants of a prospective study in which CEM was compared with MRI for the screening of women with greater than 15% lifetime risk of developing breast cancer and who were diagnosed with breast cancer during the study. The prospective studies did not address any radiomics analysis performed in the current study. CEM and MRI were performed with a mean time of 5.8 days (range, 0-66 days) within each other, and only patients who underwent 3T breast MRI at our institution were included. All patients except one (n = 49, 98%) underwent MRI and CEM within one month of each other, with a mean of 5.8 days between the two examination. In one patient, MRI and CEM were performed 66 days apart from each other.
Sixty-nine patients were excluded for the following reasons: (a) different field strength of the MRI magnet (n = 18); (b) time interval between examinations was over 10 weeks (n = 1); (c) patients receiving neoadjuvant treatment (n = 49); and (e) data not available for retrospective analysis (n = 2). Figure 1 summarizes the flowchart of patient inclusion and the lesion burden of included patients.

Contrast-Enhanced Mammography Technique
The CEM technique has been previously described [26]. CEM was performed by using a GE SenoBright mammography unit. Patients were given 1.5 mL/kg body weight of Omnipaque 350 (iohexol, GE, Millwaukee, USA) (https://www.gehealthcare.com/products/contrastmedia/omnipaque) through a 20-gauge needle, at an injection rate of 3 mL/s, with a maximum injected volume of 150 mL. Once contrast injection was complete, there was a 2.5-3 min delay during which time the patient was positioned for her mammogram. Mammographic imaging was then performed with almost simultaneous low-energy (26-30 kVp) and high-energy (45-49 kVp) images. Medio-lateral-oblique (MLO) and craniocaudal (CC) views of each breast were obtained and completed within 5 min of completion of contrast injection. The total examination time was approximately 8-9 min. The low-energy images which were generated were interpreted as the digital mammogram. A recombination algorithm subtracted out the unenhanced breast tissue and provided a subtracted image that highlighted areas of contrast enhancement.

Magnetic Resonance Imaging Technique
MRI examinations were performed with the patient prone on a 3.0-T commercially available system (General Electric Medical Systems, Milwaukee, WI, USA), using a dedicated 8-channel or 16channel phase array breast coil. Integrated Parallel Acquisition Techniques (iPAT) were utilized for imaging both breasts simultaneously. The standard MRI examination included a localizing sequence followed by sagittal fat-suppressed and sagittal T1-weighted sequences.
Two DCE-MRI protocols were used for the post-contrast phases, sagittal and axial. In Table 1, DCE-MRI parameters for both sagittal and axial protocols are summarized. Unenhanced images were subtracted from the contrast-enhanced images on a pixel-by-pixel basis, producing subtracted postcontrast subtraction sequences. Maximum-intensity projection images were generated by utilizing the first post-contrast and the first post-contrast subtracted data. Gadoteratemeglumine (Gd-DOTA;

Contrast-Enhanced Mammography Technique
The CEM technique has been previously described [26]. CEM was performed by using a GE SenoBright mammography unit. Patients were given 1.5 mL/kg body weight of Omnipaque 350 (iohexol, GE, Millwaukee, USA) (https://www.gehealthcare.com/products/contrast-media/omnipaque) through a 20-gauge needle, at an injection rate of 3 mL/s, with a maximum injected volume of 150 mL. Once contrast injection was complete, there was a 2.5-3 min delay during which time the patient was positioned for her mammogram. Mammographic imaging was then performed with almost simultaneous low-energy (26-30 kVp) and high-energy (45-49 kVp) images. Medio-lateral-oblique (MLO) and craniocaudal (CC) views of each breast were obtained and completed within 5 min of completion of contrast injection. The total examination time was approximately 8-9 min. The low-energy images which were generated were interpreted as the digital mammogram. A recombination algorithm subtracted out the unenhanced breast tissue and provided a subtracted image that highlighted areas of contrast enhancement.

Magnetic Resonance Imaging Technique
MRI examinations were performed with the patient prone on a 3.0-T commercially available system (General Electric Medical Systems, Milwaukee, WI, USA), using a dedicated 8-channel or 16-channel phase array breast coil. Integrated Parallel Acquisition Techniques (iPAT) were utilized for imaging both breasts simultaneously. The standard MRI examination included a localizing sequence followed by sagittal fat-suppressed and sagittal T1-weighted sequences.
Two DCE-MRI protocols were used for the post-contrast phases, sagittal and axial. In Table 1, DCE-MRI parameters for both sagittal and axial protocols are summarized. Unenhanced images were subtracted from the contrast-enhanced images on a pixel-by-pixel basis, producing subtracted post-contrast subtraction sequences. Maximum-intensity projection images were generated by utilizing the first post-contrast and the first post-contrast subtracted data. Gadoteratemeglumine (Gd-DOTA; Diagnostics 2020, 10, 492 4 of 11 Dotarem ® , Guerbet, Roissy, France) was injected intravenously as a bolus (0.1 mmol/kg body weight), by a power injector (Spectris Solaris EP, Medrad, Pittsburgh, PA, USA), at 4 mL/s, followed by a 20 mL saline flush.

Segmentation of Lesions
All DICOM (Digital Imaging and Communications in Medicine) images were transferred to a database and loaded into the open-source image processing tool OsiriX (OsiriX Foundation, Geneva, Switzerland). Lesion segmentation was performed on a per-patient basis, and in patients with multifocal disease, the lesion with the largest diameter was selected (Figures 2 and 3). For texture analysis, a fellowship-trained radiologist with over 4 years of experience in breast imaging manually delineated the borders of the lesions for both CEM and DCE-MRI DICOM images, including all the enhancing parts of the index cancer, excluding foreign bodies (i.e., clips). For CEM DICOM images, the borders were delineated in either the CC or MLO view, depending on which view the tumor was most conspicuous. Radiomics analysis was performed by using MaZda software (Technical University of Lodz, Poland). Radiomic features were derived from the first-order histogram (HIS): co-occurrence matrix (COM), run-length matrix (RLM), absolute gradient (GRA), autoregressive model (ARM), the discrete Haar wavelet transform (WAV), and lesion geometry (GEO). In Supplementary Materials Table S1, all the specifics of the used radiomics features are listed. For further information about the radiomic features used, please refer to Szczypiński et al. [27].
view the tumor was most conspicuous. Radiomics analysis was performed by using MaZda software (Technical University of Lodz, Poland). Radiomic features were derived from the first-order histogram (HIS): co-occurrence matrix (COM), run-length matrix (RLM), absolute gradient (GRA), autoregressive model (ARM), the discrete Haar wavelet transform (WAV), and lesion geometry (GEO). In Supplementary Materials Table S1, all the specifics of the used radiomics features are listed. For further information about the radiomic features used, please refer to Szczypiński et al. [27].  The breast is heterogeneously dense with moderate background enhancement. In the retroareolar right breast, there is a 22 mm spiculated mass containing a localizing clip from biopsy. The mass is inseparable from the nipple, which appears slightly retracted, and there is associated skin thickening. The lesion was manually segmented (2D red region of interest).

Histopathology
Histopathology results, including grade and IHC status, were recorded from biopsy specimens and served as the reference standard. An Automated Ventana Benchmark XT device was used as standard protocol to evaluate estrogen receptor (ER), progesterone receptor (PR), and human The breast is heterogeneously dense with moderate background enhancement. In the retroareolar right breast, there is a 22 mm spiculated mass containing a localizing clip from biopsy. The mass is inseparable from the nipple, which appears slightly retracted, and there is associated skin thickening. The lesion was manually segmented (2D red region of interest).

Histopathology
Histopathology results, including grade and IHC status, were recorded from biopsy specimens and served as the reference standard. An Automated Ventana Benchmark XT device was used as standard protocol to evaluate estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) status according to current American Society of Clinical Oncology/United States and Canadian Academy of Pathology guidelines, with 1% staining being considered positive [28]. HER2 status of 0 or 1+ was considered negative, 2+ equivocal, and 3+ positive. Patients with equivocal HER2 were evaluated, using fluorescence in situ hybridization, and considered as positive if HER2 gene amplification was observed [29].

Statistical Analysis
Fisher, probability of error and average correlation (POE + ACC), and mutual information (MI) coefficients were used for feature selection. Linear discriminant analysis, followed by k-nearest neighbor classification (with leave-one-out cross-validation), was used for pairwise texture-based separation of tumor invasiveness/hormonal status/tumor grade. Radiomics parameters were correlated with tumor invasiveness according to histology (invasive vs. non-invasive), hormonal status (HR+ vs. HR−), and grading (low grade G1 + G2 vs. high grade G3). Inter-rater agreement of MRI fibroglandular tissue (FGT), and background parenchymal enhancement (BPE) in both CEM and MRI was assessed by using kappa statistics with ≤0 as indicating no agreement, 0.01-0.20 as none to slight, 0.21-0.40 as fair, 0.41-0.60 as moderate, 0.61-0.80 as substantial, and 0.81-1.00 as almost perfect agreement.

Patient and Breast Cancer Characteristics
A total of 49 histopathologically verified breast cancers (mean size 24.7 ± 20 mm; range 4-87 mm) in 48 women (mean age, 50.7 ± 8 years; range, 38-74 years) were detected with CEM and MRI.

Imaging Modality
Cancer Subtype

Inter-Rater Agreement
There was moderate inter-rater agreement for density assessed by using the BI-RADS density categories (k = 0.652). Substantial inter-rater agreement was reached (k = 0.738) for the assessment of fibroglandular tissue on breast MRI. BPE was calculated for both CEM and MRI. Substantial agreement was found for BPE on CEM (k = 0.629) and on MRI (k = 0.789).

Discussion
In breast cancer, relevant prognostic and predictive factors include receptor status, tumor grade, and invasiveness, which are usually assessed with invasive tissue sampling. However, a single biopsy Diagnostics 2020, 10, 492 8 of 11 of a potentially heterogenous tumor can provide a mere snapshot of tumor biology and is subject to site selection and sampling bias; thus, it may not be representative of tumor biology in its entirety. Therefore, there is a strong argument for the development of alternative and more comprehensive means to provide prognostic and predictive information that is derived from the tumor in its entirety. In this context, there is a unique opportunity for breast imaging coupled with radiomics analysis.
In this feasibility study that performed an intra-individual comparison of CEM and DCE-MRI, results indicate the potential of both CEM and DCE-MRI for the non-invasive differentiation of invasive and non-invasive breast cancers, HR+ and HR− breast cancers, and low-grade and high-grade breast tumors with similar high accuracies.
When conducting a radiomics analysis, Bhooshan et al. [30] found in their study that 353 MRI cases that DCE-MRI radiomics features were able to distinguish between non-invasive and invasive breast cancer types. In a radiomics analysis, Bhooshan et al. [31] also found that DCE-MRI was promising for distinguishing breast tumors of different grades. Different studies have reported significant correlations between imaging-enhancement kinetics from DCE-MRI and molecular breast cancer subtypes [22,[32][33][34]. Mazurowski et al. [34] performed radiogenomic analysis on a subset of 48 cases and found an increased ratio of tumor to background parenchymal enhancement in the luminal B subtype, compared with other subtypes. Yamaguchi et al. [22] investigated the relationship between the kinetic curve pattern and molecular subtypes and showed that HR+/HER2− and triple-negative (TN) breast cancer subtypes demonstrated less washout on the delayed phase of enhancement compared with HER2-luminal (HR+/HER2+) and HER2-enriched (HR−/HER2+) subtypes. Blaschke et al. [33] reported that HER2+ tumors have a more rapid initial phase enhancement compared with other molecular subtypes. Li et al. [32] observed the faster contrast uptake in ER/PR−, and TN cancers relative to ER/PR+ and non-TN cancers. All of these studies indicate that enhancement kinetics with DCE-MRI may reflect underlying tumor biological characteristics.
To date, the evidence for radiomics using CEM is scarce, with only few published studies. Patel et al. [24] evaluated the use of a computer-aided diagnosis (CAD)-CEM in the diagnostic performance of CEM, compared with that of experienced radiologists. The authors constructed a predictive model by using a support vector machine (SVM) classification method with a set of both morphologic and textural features extracted from the low-energy and recombined images of 50 lesions. Based on the SVM classification, CAD-CEM achieved an overall diagnostic accuracy of 90% that outdid the diagnostic accuracy of the two radiologists, respectively of 78% and 86%. They therefore concluded that CAD-CEDM can provide complementary information to radiologists, mainly by reducing the number of false-positive findings. In a different study from the same group, Danala et al. [23] used the CAD scheme of CEDM images to classify breast masses. The authors used the segmentation results obtained from dual-energy subtracted images of a CEDM dataset of 111 breast lesions to build a multilayer perceptron machine learning system able to classify mass lesions. The use of CAD-CEDM in their study increased the accuracy in mass region segmentation.
While there is now a growing body of evidence that MRI and radiomics can add valuable information on tumor receptor status, grade, and molecular subtypes, data for CEM in this context are scarce. CEM is an emerging technology in breast imaging that, like MRI, combines a morphologic evaluation of the breast with an assessment of tumor neovascularity as an indicator of malignancy. CEM might therefore serve an alternative for patients who cannot undergo MRI. MR imaging is expensive and time-consuming and cannot be performed on all patients. Women who are claustrophobic and women with pacemakers or other implanted metallic materials cannot undergo breast MRI. In analogy to breast MRI, CEM is able to detect enhancing tumor vascularity after contrast administration with a sensitivity in detecting breast cancer of 96-100% [17]. Our group has recently demonstrated that there is a potential role for CEM and radiomics in the non-invasive differentiation of tumors with different degrees of invasiveness, hormone receptor status, and tumor grade [25], yet the question of whether the results are comparable to the more established technique, in this context, that is MRI, remained unanswered. In the current study, we investigated and performed an intra-individual comparison of the potential of radiomics analysis of CEM and DCE-MRI of the breast for the non-invasive differentiation of tumor invasiveness, hormone receptor status, and tumor grade in patients with primary breast cancer. Initial results indicate a potential for both techniques, with similar diagnostic accuracies.
We acknowledge some limitations of this study. The sample size is relatively small. It has to be noted that, in particular, that the number of HER2-positive breast cancers, with four cases, and pure non-invasive cancers, with four cases, is small; therefore, results, in this respect, are preliminary and have to be confirmed in larger studies. In addition, due to the small sample size, we did not divide our study population into a training dataset and a validation dataset, but used a k-nearest neighbor classifier with leave-one-out cross-validation, which does not require two separate datasets and which has been used successfully in numerous other studies [35][36][37]. Inherent to CEM technology, we could evaluate only tumor radiomics features from 2D images, which might not be as representative of tumor heterogeneity as a 3D volumetric assessment with DCE-MRI. However, despite this limitation, we achieved good results for the differentiation of tumors with different tumor receptor status, as well as invasiveness, indicating that valuable information can also be derived from 2D image analysis. All CEM and MRI examinations were performed with the same machine, at the same field strength, which assures standardized image acquisition, yet our results might not be equally applicable across vendor platforms.
In conclusion, preliminary results of this intra-individual comparison study indicate a potential of both radiomics analysis of DCE-MRI and CEM for non-invasive assessment of tumor-invasiveness, hormone receptor status, and tumor grade. CEM can therefore serve as an alternative when MRI is not available or there are contraindications for MRI and/or MRI contrast agents.