Relationships Between Human-Extracted MRI Tumor Phenotypes of Breast Cancer and Clinical Prognostic Indicators Including Receptor Status and Molecular Subtype

Abstract

PURPOSE:

The purpose of this study was to investigate if human-extracted MRI tumor phenotypes of breast cancer could predict receptor status and tumor molecular subtype using MRIs from The Cancer Genome Atlas project.

MATERIALS AND METHODS:

Our retrospective interpretation study utilized the analysis of HIPAA-compliant breast MRI data from The Cancer Imaging Archive. One hundred and seven preoperative breast MRIs of biopsy proven invasive breast cancers were analyzed by 3 fellowship-trained breast-imaging radiologists. Each study was scored according to the Breast Imaging Reporting and Data System lexicon for mass and nonmass features. The Spearman rank correlation was used for association analysis of continuous variables; the Kruskal-Wallis test was used for associating continuous outcomes with categorical variables. The Fisher-exact test was used to assess correlations between categorical image-derived features and receptor status. Prediction of estrogen receptor (ER), progesterone receptor, human epidermal growth factor receptor, and molecular subtype were performed using random forest classifiers.

RESULTS:

ER+ tumors were associated with the absence of rim enhancement (P = 0.019, odds ratio [OR] 5.5), heterogeneous internal enhancement (P = 0.02, OR 6.5), peritumoral edema (P = 0.0001, OR 10.0), and axillary adenopathy (P = 0.04, OR 4.4). ER+ tumors were smaller than ER− tumors (23.7 mm vs 29.2 mm, P = 0.02, OR 8.2). All of these variables except the lack of axillary adenopathy were also associated with progesterone receptor+ status. Luminal A tumors (n = 57) were smaller compared to nonLuminal A (21.8 mm vs 27.5 mm, P = 0.035, OR 7.3) and lacked peritumoral edema (P = 0.001, OR 6.8). Basal like tumors were associated with heterogeneous internal enhancement (P = 0.05, OR 10.1), rim enhancement (P = 0.05, OR6.9), and perituomral edema (P = 0.0001, OR 13.8).

CONCLUSIONS:

Human extracted MRI tumor phenotypes may be able to differentiate those tumors with a more favorable clinical prognosis from their more aggressive counterparts.

Introduction

Breast cancers are heterogeneous with different molecular profiles, clinical behaviors, responses to therapy, and outcomes. The diverse spectrum of breast cancers has prompted extensive research assessing the underpinnings of this complex disease. The importance of tumor biology impacting prognosis and directing treatment has spurred work to identify biomarkers in an effort to predict tumor biology.^1–4 Several commercially available tumor assays are currently in utilized by clinicians treating breast cancer in order to best tailor specific treatments for differing tumor subtypes based on differing tumor biology, such as the PAM50 assay. The PAM 50 assay analyzes the activity of 58 genes to estimate the risk of distant recurrence of hormone receptor positive breast cancer up to 10 years after diagnosis.⁵ Comparing gene expression profiles from each patient’s tumor with the 4 PAM50 molecular profiles is used to identify clinically relevant molecular subtypes of breast cancer-Luminal A, Luminal B, HER2 enriched, and basal like (BL).⁵ The spectrum of molecular subtypes confers different patterns of disease expression, response to treatment, and prognosis.^1–4,6

At present, MRI of the breast remains a sensitive technique for detecting and characterizing breast masses given its ability to provide both spatial and kinetic information. MRI is particularly reliable in detailing tumor volume, extent, associated lymphadenopathy, and involvement of the chest wall, skin, and the nipple-areolar complex.^7,8 A recent trend toward identifying surrogate markers for genetic testing has resulted in numerous publications describing relationships between MRI based phenotypes and tumor genetics.^9–13 Several groups have reported primarily on computer extracted imaging features and their ability to predict tumor genetics.^{9,11,12,14–34} Yamaguchi et al described the relative lack of contrast washout with Luminal A tumors compared to non-Luminal A tumors which they postulated was due to differential histological components (ie, Luminal A tumors were more likely to have associated ductal carcinoma in situ resulting in less washout).¹⁵ Blaschke et al described an association between rapid contrast uptake and rapid washout with Luminal B and human epidermal growth factor receptor (HER2) subtype cancers which they postulated could be due to the associations of these tumors with vascular endothelial growth factor and the propensity for neoangiogenesis and therefore early contrast enhancement.¹⁶

Grimm et al found that differing molecular subtypes had different disease distribution patterns (ie, extent of disease) on MRI with HER2 and Luminal B tumors associated with a higher likelihood of multifocal and/or multicentric disease and nodal metastases.¹⁷ Li et al described associations between computer-extracted MRI phenotypes and hormone receptor status and found that estrogen receptor (ER) negative tumors tended to be larger, more irregular, more heterogeneous in contrast uptake, and have faster contrast uptake than ER+ tumors.¹² The results of these studies offer a possible framework in which to further understand image based phenotypes or biomarkers of clinically relevant histopathologic prognostic indicators.

In an effort to assess whether image-based phenotypes are associated with specific tumor genetics, and therefore able to predict tumor aggressiveness, The Cancer Imaging Archive (TCIA) was formed through multi-institutional participation providing breast imaging studies from those genetically fully sequenced tumors in The Cancer Genome Atlas (TCGA) database.^35,36 Utilizing the TCIA database, we sought to assess whether human-extracted MRI features of known cancers in the TCGA database were associated with tumor receptor status (ie, ER, Progesterone Receptor [PR], and HER2) and the molecular subtypes of breast cancer based on the PAM50 assay.

Methods

Study Population

De-identified patient data were downloaded from the National Cancer Institute’s TCGA Breast Invasive Carcinoma and TCIA initiatives under IRB-approved HIPAA compliant protocols, this study was deemed exempt from institutional IRB approval and informed consent was waived. At the time of our study, 108 cases with breast MRI data, a subset of the entire TCGA breast cancer cases, had been collected and made available in the TCIA (http://www.cancerimagingarchive.net).³⁵ One patient was excluded due to incomplete image sequences, leaving the total number of patients included in the study at 107 (Table 1), “entire study population.” Those patients with “research based” PAM50 assay results and complete MRI data available (“PAM 50 subgroup,” n = 88) were included in a subgroup analysis of imaging phenotype predictors of molecular subtypes (Table 2). Of note, for these TCGA samples, the clinical PAM50 assay results could not be obtained—only the “research-based” assay outputs as described by Parker et al were available.⁵ Cancer subtypes were determined by using the PAM50 classifier on TCGA mRNAseq data as described by Cirello et al.³⁷ 84–91 patients have been previously reported.^9–13 The prior articles reported on the ability of computer extracted MR imaging phenotypes of the known breast cancers in the common database to predict stage, molecular classifications, and recurrence scores of the known breast cancers in the current study.^9–13 Our study included all patients with available MRI and TCGA clinical data and included all patients with available PAM50 assays. Additionally, the current study focused exclusively on the ability of human extracted or radiologist scored MRI phenotypes to predict receptor status and molecular subtypes of breast cancer.

TABLE 1.

Patient demographics, entire study population (n = 107)

Patient characteristics
Sex	Women
Age (mean, range)	53.7, (29–82)
Tumor characteristics
Size (mean, range)	24.6 mm, (7.8–87.6 mm)
Origin (frequency)
Ductal	91 (85%)
Lobular	12 (11.2%)
Mixed	4 (3.7%)
ER Status
Positive	90 (84.1%)
Negative	17 (15.9%)
PR Status
Positive	81 (75.7%)
Negative	26 (24.3%)
HER2 Status
Positive	22 (20.5%)
Negative	85 (79.4%)

Abbreviations: ER, estrogen receptor; PR, progesterone receptor, HER2, human epidermal growth factor receptor.

TABLE 2.

PAM50 tumor subtypes, PAM50 subgroup (n = 88)

Luminal A	57 (64.8%)
Luminal B	10(11.4%)
HER2 Enriched	5 (5.7%)
Basal Like	11 (12.5%)
Normal Like	5 (5.7%)

Image Data

All MRIs were acquired utilizing a standard double breast coil and were performed on 1.5 T GE (Milwaukee, WI) whole body MRI systems. T1-weighted dynamic contrast-enhanced and T2-weighted images were analyzed for this study, including 1 pre and 3–5 postcontrast images obtained using a T1-weighted 3 dimensional spoiled gradient echo sequence with a gadolinium-based contrast agent. In-plane resolution on the dynamic series was 0.53–0.86 mm, and spacing between slices was 2–3 mm. Imaging was performed across 4 institutions including Memorial Sloan Kettering Cancer Center, Mayo Clinic, University of Pittsburgh Medical Center, and Roswell Park Cancer Institute. All 107 TCGA patients with biopsy proven invasive breast cancer and available breast MRI results were assessed by a panel of radiologists (TCGA Breast Phenotype Research Group) using Clear Canvas software (ClearCanvas, Toronto, Ontario, Canada).³⁸ Imaging was scored based on a variety of mass and nonmass features according to the Breast Imaging Reporting and Data System 2013 edition³⁹ (Table 3). Each breast MRI examination was reviewed and annotated independently by 3 of the 11 expert board-certified breast radiologists who were blinded to outcomes with final scoring reached by consensus. The breast imaging experience of radiologists in this study ranged from 4 to 29 years.

TABLE 3.

Imaging phenotypes (n = 107)—distribution of radiologist scored BIRADS features in the entire study population

Breast observations	Radiologist score (frequency)
Breast composition	Extreme (13)/Heterogeneous (43)/Scattered (42)/Fatty (9)
Background enhancement	Marked (5)/Moderate (23)/Mild (53)/Minimal (26)
Extent	Unicentric (58)/Multicentric (21)/Multifocal (28)
Mass observations
Shape	Round-oval (38)/Irregular (69)
Margin	Circumscribed (8)/Spiculated (67)/Irregular (32)
T2 pattern	Homogeneous (56)/Heterogeneous (51)
T2 signal intensity	Hyperintense (25)/Isointense (78)/Hypointense (4)
T2 peritumoral edema	Present (29)/Absent (78)
Dark internal septations	Present (2)/Absent (105)
Internal enhancement	Heterogeneous (63)/Homogeneous (44)
Rim enhancement	Present (17)/Absent (90)
Lesion size	Maximum diameter
Axillary adenopathy	Present (27)/Absent (80)
Pectoralis invasion	Present (2)/Absent (105)
Chest wall invasion	Present (0)/Absent (107)
Edema	Present (1)/Absent (0)
Nipple Inversion	Present/Absent
Nipple retraction	Present (6)/Absent (101)
Nipple Invastion	Present (3)/Absent (104)
Skin thickening	Present/Absent
Skin invasion	Present (2)/Absent (105)
Skin retraction	Present/Absent
Nonmass observations
Associated nonmass enhancement	Present (33)/Absent (74)
Nonmass distribution	Focal (16)/Linear (16)/Diffuse (11)/Segmental (0)/Regional (16)/Multiple Regions (19)
Nonmass internal enhancement	Heterogeneous (28)/Clumped (44)/Clustered Ring (14)

Statistical Analysis

Batch corrected gene expression data was obtained from the TCGA Data Portal. The Spearman rank correlation was used for association analysis of continuous variables; the Kruskal-Wallis test was used for associating continuous outcomes with categorical variables.⁴⁰ The Fisher-exact test was used to assess correlations between categorical image-derived features and ER, PR, and HER2 receptor status.⁴¹ Correction for multiple hypothesis testing was conducted using the Benjamini-Hochberg method to control the False Discovery Rate.⁴² Associations with adjusted P-values no larger than 0.05 were deemed as statistically significant and selected for interpretation. Prediction of ER, PR, HER2, and molecular subtype (from PAM50 data) were performed using random forest classifiers. The random forest classifier is an ensemble of decision-trees, can handle a mixture of numeric (eg, lesion size) and categorical data (eg, Breast Imaging Reporting and Data System margin descriptor) to predict class labels, and is capable of handling large numbers of features. The prediction performance is assessed via a bootstrap sampling (using an “out-of-bag” set for model assessment).⁴³ Variable importance plots were obtained by measuring the improvement in the probability of correct classification induced by a specific variable.⁴⁴ An adjusted P-value of 0.05 was selected as the threshold for determining statistical significance. For ER and/or PR status prediction, the difference of the random forest classifier’s performance using area under the receiver operating characteristic curve, area under the curve (AUC) relative to random classi-fication (AUC = 0.5) is assessed via P-value from a Mann-Whitney hypothesis test (using R-package, “verification”).⁴⁵

Results

The entire study population of 107 biopsy proven invasive breast cancers included 91(85%) ductal, 12 (11.2%) lobular, and 4 (3.7%) mixed tumors, the majority of which were ER+ (n = 90). Thirteen cases were triple negative breast cancers (TNBCs) (Table 1). Of those cases with PAM50 assay data available, the PAM50 subgroup, (n = 88), 76 (86.4%) were ductal, 10 (11.4%) were lobular, and 2 (2.2%) were mixed, the majority of which, 64.8%, were classified as Luminal A (n = 57). Eleven of the PAM50 subgroup tumors were BL, of which 10 were TNBC with the final BL tumor being ER+/PR+/HER2+ (Table 2).

MRI Phenotype Association with Histopathologic Variables

Based on the entire study population (n = 107), the absence of all of the following were associated with ER+ tumors: rim enhancement (P = 0.019, odds ratio [OR] 5.5), heterogeneous internal enhancement (P = 0.02, OR 6.5), peritumoral edema (P = 0.0001, OR 10.0), and axillary adenopathy (P = 0.04, OR 4.4). ER+ tumors were smaller than ER+ tumors (23.7 mm vs 29.2 mm, P = 0.02, OR 8.2). All of these variables except the lack of axillary adenopathy also maintained their statistically significant associations with PR status (Table 4).

TABLE 4.

Imaging phenotype associations—Receptor status and molecular subtype associations with the three imaging phenotypes which emerged as main features associated with specific tumor profile. Of note, the 5 normal like tumors from the PAM50 subgroup (n = 88) were not included as no imaging phenotypes were associated with this molecular subtype

Entire study population (n = 107)	Average size (mm)	Rim enhancement (%)	Heterogeneous internal enhancement (%)	Peritumoral edema (%)
ER+	23.7	11 (10/90)	53 (48/90)	19 (17/90)
ER-	29.2	41 (7/17)	88 (15/17)	70 (12/17)
PR+	23.8	11 (9/81)	50 (41/81)	58 (15/81)
PR-	27.1	31 (8/26)	85 (22/26)	54 (14/26)
HER2+	23.5	18(4/22)	59(13/22)	32 (7/22)
HER2-	24.8	15 (13/85)	58 (50/85)	26 (22/85)
TNBC	27.9	46 (6/13)	92 (12/13)	77 (10/13)
PAM50 subgroup (n = 88)
Lum A	21.8	9 (5/57)	47 (27/57)	10 (6/57)
Lum B	25.5	10 (1/10)	70 (7/10)	20 (2/10)
HER2	27.8	40 (2/5)	40 (2/5)	60 (3/5)
Basal Like	29.2	45 (5/11)	91 (10/11)	62 (8/11)

Abbreviations: ER, estrogen receptor; PR, progesterone receptor, HER2, human epidermal growth factor receptor; TNBC, triple negative breast cancer; Lum A, Luminal A; Lum B, Luminal B.

Based on the PAM50 subgroup (n = 88), when looking at the Luminal A tumors compared to non-Luminal A (n = 57), Luminal A tumors were smaller (21.8 mm vs 27.5 mm for non-Luminal A, P = 0.035, OR 7.3) and they lacked peritumoral edema (P = 0.001, OR 6.8). When comparing BL tumors (n = 13) to non-BL (n = 75), BL tumors were associated with heterogeneous internal enhancement (P = 0.05, OR 10.1), rim enhancement (P = 0.05, OR 6.9), and perituomral edema (P = 0.0001, OR 13.8) (Table 5).

TABLE 5.

Imaging phenotype associations—associations between receptor status of tumors in entire the study population (n=107) as well as in the PAM50 subgroup and imaging phenotypes which reached statistical significance

Entire study population (n = 107)	Imaging phenotype	BH corrected P value	Odds ratio
Receptor status:
ER+	T2 Peritumoral edema absent	0.0002	10.0
	Lesion Size	0.02	8.2
	Heterogeneous internal enhancement absent	0.02	6.5
	Rim enhancement absent	0.02	5.5
	Axillary LAD absent	0.04	4.4
PR+	T2 Peritumoral edema absent	0.001	5.0
	Lesion Size	0.05	4.7
	Heterogeneous internal enhancement absent	0.02	3.9
	Rim enhancement absent	0.05	3.5
PAM50 subgroup (n=88)
Molecular subtype
Luminal A	Lesion size	0.04	7.3
	T2 peritumoral edema absent	0.001	6.8
Basal Like	T2 peritumoral edema present	0.001	13.8
	Rim enhancement	0.05	6.9

Abbreviations: ER, estrogen receptor; PR, progesterone receptor.

Imaging Phenotype Importance in Predicting Receptor Status

In the entire study population (n = 107), the following variables ranked at the top of the variable importance plots using the random forest classifier to predict receptor status (Fig 1A–B): ER status: lesion size, peritumoral edema, and internal enhancement (AUC 0.8, 95% confidence interval [CI] 0.654–0.88, P = 0.0001); PR status: lesion size, peritumoral edema, and internal enhancement (AUC 0.633, 95% CI 0.505–0.776, P = 0.02). MRI tumor phenotypes did not reliably predict HER2 status (AUC 0.48, P = 0.4). Variable importance plots rank order those features most important in predicting a given receptor status, for both ER and PR tasks, a sharp drop-off was noted after the first few features making those at the top most important in predicting receptor status in this population.

FIG 1.

(A, B) Receptor status prediction tasks and associated variable importance plots from full study population (n = 107). AUC for all MRI phenotypes used to predict receptor status and corresponding variable importance plots. The AUC for the ER status prediction task is 0.8 with confidence intervals [0.654, 0.88], and P-value 0. For PR status prediction the AUC is 0.633 with confidence interval [0.505, 0.776], P-value = 0.02. For both ER and PR status, imaging features performed well at predicting receptor status. Variable importance plots rank order those features most important in predicting a given receptor status, for both ER and PR tasks, a sharp drop-off was noted after the first few features making those at the top most important in predicting receptor status in this population.

Discussion

The results from this study demonstrate that human extracted MRI features may be able to differentiate tumors with favorable prognosis from their more aggressive counterparts. Lesion size, internal enhancement characteristics, and the presence or absence of peritu-moral edema emerged as potential predictors of tumor receptor pro-files and molecular subtypes. From our results, it is apparent that those tumors which carry a more favorable prognosis and have known better response to treatment (ie, ER+, PR+, and Luminal A tumors) tended to be smaller, lacked heterogeneous internal enhancement, rim enhancement, peritumoral edema, and, with ER+ tumors, lacked axillary adenopathy. Those tumors which demonstrated heterogeneous internal enhancement, rim enhancement, and peritumoral edema conferred a worse prognosis as they were associated with triple negative and BL tumors which have a known resistance to treatment and poorer outcomes.^46–48

The relationship between imaging phenotypes and tumor biology has been suggested in recent publications assessing breast MRI phenotypes and tumor genetics. Several published studies describe rim enhancement and heterogeneous internal enhancement associated with TNBC.^49–53 Given the known crossover between TNBC and BL subtypes, our results suggest that these features when present may indicate the more aggressive TNBC and BL subtypes. Authors have hypothesized these features may be related to differential greater microvessel density, expression levels of vascular endothelial growth factor, and overall higher cellular proliferation markers leading to larger tumors with heterogeneous enhancement patterns.^49–53 Lesion size was also an important variable in predicting receptor status and ranked at or near the top of all variable importance plots for each receptor subtype, with ER+ and PR+ tumors tending to be smaller than their receptor negative counterparts (Table 4, Fig 1A–B). These results are in keeping with recent publications highlighting the importance of size in predicting overall pathologic stage,⁹ risk of recurrence,¹¹ and in the aforementioned reports of the tendency of more aggressive BL and TNBC subtypes to be larger than their Luminal A/B counterparts. In Burnside and Drukker et al’s study evaluating computer extracted image phenotypes from breast MRI and their ability to predict stage, size was the most powerful predictor of overall pathologic stage.⁹ Li et al recently reported the ability of computer extracted imaging features to predict risk of recurrence and found that heterogeneous enhancement (enhancement texture) and larger size correlated with a higher risk of recurrence based on various gene assay models, including PAM 50.¹¹

There are several limitations of this study. First, the relatively small sample size resulted from the relatively limited number of patients with breast MRIs in the TCGA and/or TCIA databases with even fewer having PAM 50 assay data available. Given that the majority of tumors in this dataset were ER+, PR+, HER2−, and Luminal A, the receptor and molecular subtype analyses may have been skewed given relative fewer numbers of receptor negative and nonLuminal A tumors. The inability to identify any meaningful associations between image phenotypes and HER2 status is likely due to the relative small number of HER2+ tumors in our dataset (n = 22). Future studies will hopefully allow for a data set with more balanced molecular and receptor subtypes. The MRIs included in this study were acquired more than 10 years ago at 4 different institutions, each with variations in image acquisition including differing equipment and protocols, and therefore may not be representative of current MR practice standards with newer advanced technology and protocols. Despite such limitations, the TCGA and/or TCIA databases remain the largest publicly available resource for radiogenomics research. Future directions include expanding the study to a larger data set, including more current imaging technology and standardized protocols, with hopes to identify reproducible imaging biomarkers of clinically relevant prognostic indicators.

Conclusions

Human extracted MRI-based tumor phenotypes may be able to differentiate those tumors with a more favorable clinical prognosis from their more aggressive counterparts. The use of such MRI-based tumor phenotypes has the potential to provide useful prognostic information, and lead to a better understanding of tumor biology and aid in tailoring treatments for more precise breast cancer care.