Radiomic signatures with contrast-enhanced magnetic resonance imaging for the assessment of breast cancer receptor status and molecular subtypes: initial results

Breast cancer is a heterogeneous disease with varying clinical presentations, subtypes, and treatment responses [1‐7]. Although traditional prognostic and predictive factors such as tumor size, grade, histopathologic type, and immunohistochemical receptor status are well established, it has become evident that these traditional classifications cannot fully capture the heterogeneity of breast cancer so that the classical approach of stratifying patients into treatment groups based on phenotypic biomarkers may be insufficient in some patients. In this era of precision medicine, treatments are selected based on genetic tumor characteristics. In breast cancer, gene expression profiling has revealed four main intrinsic molecular subtypes that show pervasive differences in their gene expression patterns: luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, and triple negative [8‐10]. These intrinsic molecular subtypes have different phenotypic presentations, prognosis, treatment responses, and recurrence-free and disease-specific survival leading to molecular subtype-based recommendations for systemic therapy [1‐3, 11, 12].

Assessment of molecular subtypes is currently based on either gene expression profiling or immunohistochemical (IHC) surrogates from invasive tissue sampling [10, 13, 14]. It must be noted that this approach has limitations. A single biopsy of a potentially heterogenous tumor can capture only a snapshot of the tumor tissue that is subject to selection bias and may not be completely representative of the genetic, epigenetic, and/or phenotypic alterations of the entire tumor. In addition, cancer biology is subject to change over time as well with treatment from a stem-like, therapy-resistant and a differentiated drug-sensitive phenotype in a process linked to epithelial-mesenchymal transition [15]. Therefore, there is a strong argument for the development of alternative means to derive tumor characteristics that are prognostic indicators, i.e., receptor status and molecular subtypes, from the tumor in its entirety and to spatio-longitudinal monitor tumor biology changes during treatment. This is a unique opportunity for medical imaging, and in this context, contrast-enhanced magnetic resonance imaging (CE-MRI) coupled with radiomic analyses has yielded initial encouraging results. Prior studies have investigated radiomic signatures in the breast, but the generalization of these results is limited due to utilization of different MRI protocols and scanners [11, 16‐18], results may be suboptimal due to inclusion of only a few radiomic features [11, 17‐20], and only certain subgroups of breast cancers have been investigated [21].

We hypothesized that through the extraction of radiomic signatures from standardized CE-MRI data, an accurate assessment of molecular subtypes and receptor status of breast cancers will be provided. Thus, the aim of our study was to evaluate the diagnostic performance of CE-MRI coupled with radiomic analysis for the differentiation of breast cancers of different receptor status and molecular subtype.

This retrospective study was compliant with Health Insurance Portability and Accountability Act guidelines and approved by the Institutional Review Board with a waiver of written informed consent.

Mean lesion size of the 91 treatment-naïve, biopsy-proven breast cancers in 91 patients (mean age, 48 ± 9.7 years, range, 27–68 years) was 3.5 ± 2.3 cm (range, 1–16.6 cm). There were 70 mass lesions and 21 non-mass enhancing lesions. Hormone receptor positivity was observed in 57 patients (62.6%). Forty-nine breast cancers were luminal A (53.8%), eight were luminal B (8.8%), 11 were HER2-enriched (12.1%), and 23 were triple negative (25.3%).

Results of group-wise radiomic-based classifications including numbers of features are listed in Tables 1 and 2. Detailed results of all radiomic feature-based cancer classifications including all selection algorithms and respective MDFs are given in Additional file 1: Table S1. With the exception of HER2-enriched vs. luminal A (MDF = 2), the previously selected texture features were reduced to a single MDF. Selected feature sets for classifications with accuracies > 80% are given in Table 3, while feature sets for all pairwise classifications and selection algorithms are provided in Additional file 2: Table S2. Classification accuracies ≥ 80% were considered as clinically relevant and were achieved for molecular subtypes and receptor status as listed below.

In this study, we evaluated the diagnostic performance of CE-MRI coupled with radiomic analysis for the non-invasive differentiation of breast cancers with different receptor status and molecular subtypes. We hypothesized that microstructural differences between breast cancers of different receptor status, and hence molecular subtypes, would lead to different, larger-scale gray-level patterns in CE-MR images which can be assessed quantitatively by radiomic analysis. Our results demonstrate that radiomic features derived from CE-MRI data enable the determination of receptor status and molecular subtypes with a high accuracy. CE-MRI radiomic signatures may therefore provide valuable prognostic indicators based on the entire tumor and thus may be used to monitor spatio-longitudinal tumor biology changes during treatment.

In our study, the best results were achieved for the differentiation of luminal A vs. luminal B (accuracy 84.2%), luminal B vs. triple negative (83.9%), luminal B vs. all others (89%), and HER2-enriched vs. all other cancers (81.3%). In these four pairwise comparisons, especially COM and GEO features seemed to be of importance, which highlight the relevance of structural heterogeneity on the one hand and lesion shape on the other hand. External validation of part of our results yielded satisfactory results for the differentiation of luminal A vs. luminal B (79.4%) and luminal B vs. triple negative (77.1%). Our results indicate that, since tumor biology may change in course of treatment, radiomic features may possibly be able to capture these changes noninvasively, provided that our results are verified in future studies that use larger datasets and a prospective design.

Li et al. achieved excellent results for radiomic-based separation between estrogen receptor-positive and estrogen receptor-negative cancers with an area under the curve of 0.89 [17]. Given the fact that in clinical decision-making, a distinction of positivity of estrogen receptor and progesterone receptor is not relevant, we considered either estrogen receptor or progesterone receptor positivity as hormone receptor positive and did not analyze these receptors separately. In our study, in contrast to the differentiation of luminal A/B cancers, the differentiation of hormone receptor-positive and hormone receptor-negative cancers had limited success with an accuracy of 68.1%. However, we showed excellent results for the differentiation of luminal B vs. all other subtypes (89%) and luminal A vs. B cancers (84.2%). In addition to being hormone receptor positive, luminal B cancers are also characterized by other biologic features such as higher proliferation rates and/or HER2 positivity, and it seems that this can be captured through distinctive radiomic features. All hormone receptor-positive cancers are treated with endocrine therapy alone or with chemotherapy, while luminal B cancers in contrast to luminal A cancers derive benefit from additional cytotoxic and targeted treatment. In this context, the ability to identify and non-invasively spatio-longitudinally monitor these breast cancers during treatment is particularly relevant.

In this study, we used a combination of multiple radiomic features derived from different categories (e.g., COM, RLM, and ARM), all of which capture different aspects of image texture and thus might have contributed to the excellent results. This approach differs from the majority of prior MRI radiomic studies in the field of breast cancer, in which typically very few and often COM-based radiomic features are calculated [18‐20, 32]. Sutton et al. achieved accuracies of up to 89.2% for the differentiation of molecular subtypes but used a combination of pathology-derived features and COM radiomic features [32]. In this study, we intentionally chose to rely solely on radiomic features from CE-MRI and to dispense with features derived from invasive tissue sampling. Holli-Helenius aimed to differentiate between luminal A and B subtypes in a small collective of 27 patients, using COM features for radiomic analysis [33]. With areas under the curve of 0.83–0.88, these results were very similar to those in our own study (84.2% accuracy); in accordance to Holli-Helenius, COM features were of importance for the differentiation between those two groups. Notably, the results of Holli-Helenius mainly relied on features extracted from pre-contrast MR images, which were not included in our study. Nevertheless, our results generally confirm that luminal A and B cancers can be differentiated by MRI-based radiomic features with a high level of confidence.

Wang et al. developed a model to exclusively distinguish triple-negative cancer from other subtypes and achieved an area under the curve of 0.88 using COM features extracted from both the tumor and healthy breast parenchyma [21]. These results agree with our findings where we demonstrated a differentiation of triple negative from luminal B with an accuracy of 83.9% when employing radiomic features from the tumor itself but not from the healthy breast tissue. The good differentiation between triple-negative and luminal B cancers may be explained by the fact that the geometric features (Table 3) used for this task capture the typically round/circumscribed shape of triple negative and the irregular/spiculated shape of luminal B cancers. In addition, Fan et al. aimed to develop a model to distinguish between molecular breast cancer subtypes based on CE-MRI data of the tumor and healthy breast parenchyma [19]. They used COM and GEO radiomic features, dynamic enhancement features, and clinical parameters and achieved good to excellent discrimination of luminal A, luminal B, HER2-enriched, and triple-negative cancers from all other tumors. Subtype discrimination in this study was partially less successful than our study (e.g., for luminal B cancers) even though we did not incorporate clinical parameters, and in part more successful (e.g., for triple-negative cancers). While it must be noted that Fan et al. used a training and a validation dataset—a strategy which, especially in combination with artificial neural networks, is preferable when datasets are sufficiently large—they did not use the same set of radiomic features in the two datasets; thus, no true validation of their model was performed.

External validation of part of our results yielded satisfactory results for the differentiation of luminal A vs. luminal B (79.4%) and luminal B vs. triple negative (77.1%). These results might be attributed to the feature sets produced by Fisher coefficients in these two comparisons, because they contained only GEO features for luminal A vs. luminal B and two GEO features and one HIS feature for luminal B vs. triple negative (see Additional file 2: Table S2). Unlike texture features, GEO and HIS feature should not be sensitive to MR image acquisition parameters. As expected, due to technical differences (TR/TE, scanner model and vendor, etc.) between training and validation datasets, some of the other results were poor, indicating a partial lack of generalizability to a wider clinical setting. Notably, some were instances where class imbalance led to seemingly good results despite a lack of separability, for instance the differentiation of luminal B vs. all others (75%). In contrast to prior CE-MRI radiomic studies that used heterogeneous image datasets in terms of acquisition parameters [11, 16‐18], the MRI protocol in this study was homogeneous, which might have contributed to the excellent results. Grimm et al. evaluated associations of molecular breast cancer subtypes with imaging characteristics that included morphologic, radiomic, and dynamic enhancement features using a semi-automatic approach for lesion segmentation (i.e., a fuzzy C-Means clustering algorithm [16]. They found that enhancement characteristics and morphological features were superior to radiomic features. However, in their study, there was a high level of heterogeneity in terms of scanning equipment and pulse sequence protocols. In particular, matrix size, which, through its association with spatial resolution (and thus, pixel/voxel size), has been shown to have a major impact on feature calculations [34] and thus could be responsible for the divergent results. Similar MRI acquisition protocol heterogeneities may also have contributed to the disappointing performance of radiomic features in the study by Mazurowski et al. that investigated associations of molecular subtypes and semi-automatically extracted CE-MRI data, including COM and GEO features [11]. In contrast to prior publications that investigated molecular subtyping with radiomic analysis, a novelty and strength of the present study is our assessment of a wealth of radiomic features derived from a homogeneous state-of-the-art MRI protocol. These prior divergent results highlight the necessity for rigorous standardization across institutions, vendors, and platforms to fully explore the potential of MRI radiomics in breast cancer, especially when sample sizes are limited. On the other hand, in truly large datasets (“big data”) in which a sufficiently large number of cases for each protocol variation is available, protocol heterogeneity could present an advantage with regard to generalizability.

We acknowledge the limitations of the present study. In this study, molecular subtypes were derived not from formal genetic testing but from IHC surrogates. It has to be acknowledged that there is variable agreement between classifications via these surrogates and formal genetic testing (41–100%). However, in clinical practice, molecular subtypes and subsequent clinical decision-making are most often based on IHC surrogates. Like the vast majority of previous radiomic studies, the conclusions that can be drawn from this study are somewhat limited due to its retrospective nature and the unequal distribution of molecular subtypes. Another drawback is (in comparison to luminal breast cancers) the smaller number of HER2-enriched and triple-negative cancers in our patient collective (Fig. 3); however, this reflects the normal distribution of molecular subtypes of breast cancer in a patient population. Due to this limitation, we decided not to divide our original study population into a training dataset and a validation dataset. Instead, we used a k-nearest neighbor classifier with leave-one-out cross-validation, a technique that does not require two separate datasets and has been used in numerous studies in the field [22, 35, 36], for all pairwise comparisons, even though n-fold cross-validation may have produced better results in the larger samples. It has to be acknowledged that as each training set used in leave-one-out cross-validation only differs by one sample, the models tend to be more highly correlated and less variability is encountered when training the model. Thus, there is a possibility of underestimating the error via leave-one-out cross-validation. As the dataset is relatively small (especially for some subgroups such as HER2-enriched and luminal B cancers), there is the possibility that some of our findings could be the product of overfitting. Nevertheless, we aimed to decrease the risk of overfitting by reducing the dimensionality of feature space to 10% of the samples out of the original ten features provided by MaZda (e.g., five features for 50 samples). Validation of the promising results of this initial study with larger patient numbers is currently ongoing. We acknowledge that the mean size of breast cancers included in this study was relatively large, which could explain the high number of triple-negative cancers. However, this would be the target population for neoadjuvant treatment where radiomic analysis may be used to monitor spatio-longitudinal tumor biology changes during treatment and thereby identify a treatment-resistant phenotype. Unlike techniques such as 18F-fluoroestradiol positron emission tomography, which directly visualizes functional estrogen receptors, radiomics cannot directly capture receptor density. However, estrogen/progesterone receptor responsiveness “fuels” certain growth patterns, which may be captured by radiomic signatures. GEO features are influenced by tumor size to some degree; however, the texture feature classes that reflect homogeneity/heterogeneity of the lesions should not be significantly influenced by tumor size. For instance, COM features are calculated with a maximum interpixel distance of n = 5; since in-plane resolution was approximately 1.26 mm, feature calculation would not have been affected in tumors of 1 cm or larger. We acknowledge that a comparative analysis of images and corresponding cross-sectional histological stains would be of special interest. However, a confident co-registration is challenging, and for tumors treated with neoadjuvant chemotherapy, treatment-naïve tumor specimens are not available. Breast cancers were delineated manually and on the slice with the largest lesion diameter only. The combination of CE-MRI with diffusion-weighted imaging data might further improve the results of radiomic analysis, as different aspects of tumor biology could be captured, i.e., heterogeneity in terms of perfusion on the one hand and diffusivity/cell density on the other hand. In addition, the integration of radiomic features from healthy breast parenchyma might also be of great interest [21].

In conclusion, radiomic signatures with CE-MRI may enable the assessment of breast cancer receptor status and molecular subtypes. Our initial results indicate that CE-MRI radiomic signatures may have the potential to provide prognostic indicators derived from the tumor in its entirety; theoretically, this information could be used to monitor spatio-longitudinal tumor biology changes during treatment. However, larger prospective studies and efforts in data and protocol standardization are warranted to validate these findings before definitive conclusions with regard to the value of CE-MRI radiomic features for distinguishing between breast cancer receptor status and molecular subtypes can be drawn.