Radiomic Signatures Derived from Diffusion-Weighted Imaging for the Assessment of Breast Cancer Receptor Status and Molecular Subtypes

With the revelation that breast cancer is a genetic disease, it has become evident that traditional classifications based on tumor histology, size, grade, and receptor status cannot fully capture its characteristics. Gene expression profiling has revealed four main intrinsic molecular subtypes of breast cancer that show substantial differences in phenotypic presentation, prognosis, treatment response, and outcome [1‐5]: luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched, and triple negative (TN) [6‐8]. The assessment of molecular subtypes to guide treatment selection is currently performed by gene expression profiling or immunohistochemical surrogates from tissue samples [8‐10]. However, this approach can be limited, as biopsy can only capture a small part of a potentially heterogeneous lesion. Additionally, tumor biology may change over time and during treatment due to epithelial–mesenchymal transition [11].

The emerging field of radiomics relies on the extraction of mathematical patterns that are hidden within medical images, some of which the human eye may not be able to assess, let alone quantify; these patterns may potentially predict disease genotypes and enable spatio-longitudinal monitoring of the entire tumor. Previous radiomic studies in breast cancer have primarily focused on features derived from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and their utility for the prediction of molecular subtype [12‐15], as well as tumor histology [16], risk of recurrence [17, 18], response to chemotherapy [19], and potential to metastasize [20, 21]. While DCE-MRI is undoubtedly the most sensitive modality for the assessment of breast cancer, complementary techniques such as diffusion-weighted imaging (DWI) that provides additional functional information have been developed to increase its specificity. DWI with apparent diffusion coefficient (ADC) mapping visualizes diffusivity, which indirectly reflects cell density in solid tumors.

Nevertheless, little is known about radiomic signatures derived from DWI for breast cancer characterization: the small number of previous studies of radiomics of DWI primarily evaluated differentiation between benign and malignant lesions [22‐24] or solely used histogram features for the differentiation of TN cancers vs. other breast cancer subtypes [25]. Moreover, the actual parameters used to perform DWI radiomics are not yet well established, i.e., whether only the actual solid tumor itself should be measured or if the inclusion of edema, necrosis, and inflammation is essential for the differentiation between subgroups [26, 27].

Due to the recent controversy about the safety of gadolinium-based contrast agents, there is considerable interest to develop unenhanced MRI techniques for improved breast cancer diagnosis and characterization [28‐30]. Therefore investigation of DWI radiomic signatures for the differentiation of molecular subtypes is of interest and may have manifold applications such as unenhanced assessment of tumor heterogeneity for improved biopsy planning to better guide treatment decisions or the non-invasive tumor monitoring.

We hypothesized that the spatial heterogeneity of tissue diffusivity differs between molecular breast cancer subtypes. Therefore, the aim of this study was to evaluate the diagnostic performance of radiomic features extracted from standardized DWI data, using different approaches of tumor segmentation, for the assessment of breast cancer receptor status and molecular subtypes.

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

Of the 91 treatment-naïve, biopsy-proven breast cancers, 57 were hormone receptor positive (62.6 %). Forty-nine cancers were classified as luminal A (53.8 %), 8 as luminal B (8.8 %), 11 as HER2-enriched (12.1 %), and 23 as TN (25.3 %). There were 70 mass lesions and 21 non-mass enhancing lesions on DCE-MRI. Mean lesion size was 3.5 ± 2.3 cm (range, 1–16.6 cm). Mean patient age was 48 ± 9.7 years (range, 27–68 years).

Radiomics-based classification accuracies ≥ 80 % were considered to be clinically relevant and are listed below:

In this study, we investigated the diagnostic performance of radiomic signatures derived from DWI for the assessment of breast cancer receptor status and molecular subtypes. We hypothesized that the spatial heterogeneity of tissue diffusivity differs between molecular breast cancer subtypes and could be quantified by radiomic analysis. Our results demonstrate that DWI radiomic features enable the separation between breast cancers of different receptor status and molecular subtypes with high accuracy. Two different approaches of tumor segmentation were explored in this context. Tumor segmentation directly on the ADC map yielded better results in terms of accuracy, suggesting that the solid tumor components strongly contribute to radiomic analysis rather than the inclusion of peritumoral or necrotic tissue. Radiomic signatures derived from DWI may have the potential to enable contrast agent-free spatio-longitudinal monitoring of tumor biology before and during treatment.

One of the main findings of our study is the high accuracy in the radiomic-based discrimination of luminal B and HER2 enriched breast cancers (accuracies, 100 % and 94.7 %). This specific finding might have a direct clinical consequence, as it might prevent the incorrect exclusion of women from hormone therapy when they have heterogeneous tumors. The important information that hormone receptors are present within the tumor might get lost, as biopsy can only capture a snapshot of a potentially heterogeneous tumor, and after neoadjuvant chemotherapy, no tumor cells might be left.

In the first study that investigated the utility of DWI radiomic signatures, Xie et al. investigated DWI and DCE-MRI histogram features for the differentiation of TN from other molecular subtypes with AUCs up to 0.763 [25]. Histogram features do not provide textural information in terms of spatial relationships between the signal intensities of pixels/voxels across a region or volume of interest [35]. This may explain why in this first study that investigated the utility of DWI radiomic signatures, true radiomic feature groups (e.g., COM) performed better.

Other studies have investigated radiomic signatures derived from DWI for the differentiation of benign from malignant breast lesions. Bickelhaupt et al. recently evaluated DWI radiomic signatures for the separation between benign and malignant breast lesions [22]. Although DWI radiomics yielded better results than ADC mean alone, a highly experienced breast radiologist using DCE-MRI could not be outperformed. Another study by the same group investigated a radiomic model extracted from kurtosis DWI for lesion characterization [23]. The model reduced false positives from 66 to 20 cases at a predefined sensitivity of greater than 98 %, significantly improving specificity compared with median ADC and apparent kurtosis coefficient alone. It has to be noted that in the aforementioned study, there was heterogeneity in terms of hard- and software; however, voxel size for DWI was kept constant, which might have influenced results. Parekh, et al. attempted to separate benign from malignant breast lesions using COM features derived from feature maps of DWI, DCE-MRI, T1-weighted, and T2-weighted imaging [24]. The authors found cellular heterogeneity, evaluated using entropy on the ADC map, to be significantly different between benign and malignant lesions.

In the aforementioned studies, different approaches to derive and investigate DWI radiomic signatures were used which limits generalizability of results. It stands to reason that a prerequisite for the widespread application of radiomics in the future will be the rigorous standardization in terms of data acquisition and analysis across institutions and vendors. An alternative would be the use of deep learning neural networks, i.e., with very large heterogeneous datasets in which the classifier has sufficient training data to learn what the real biological information is. In this respect, investigation of how tumors should ideally be segmented is of interest as it remains unclear whether the ROI should include solid tumor components only or additionally cover cystic, hemorrhagic, necrotic, or inflammated regions [43]. Liu, et al. found 95 radiomic features in cervical cancer to be insensitive to ROI variation among T2-weighted images, ADC maps of b800 and b1000, with ADC b1000 features yielding a lower misclassification error [27]. In the present study, only b1000 ADC maps were chosen for radiomic analysis. Recently, Wang, et al. compared whole tumor, solid portion, and peritumoral edema ROIs for the differentiation between low-grade and high-grade gliomas [26]. They found that tumor inhomogeneity parameters performed better than the ones derived from solid areas only (P = 0.048); nevertheless, only HIS and COM features were calculated. In our study, tumor segmentation directly on ADC maps performed better for the separation of molecular breast cancer subtypes than segmentation and ROI transfer from high b-value DWI, which naturally included more non-solid areas surrounding the lesion. The ADC map represents areas of tumor with actual hindered diffusivity and thus might allow a more accurate assessment of the lesion than high b-value DWI, where T2 shine through might occur. Our results suggest that solid tumor components truly contribute to radiomic analysis; however, further studies with three-dimensional evaluation of DWI radiomics in larger patient cohorts are warranted to confirm our findings.

While DWI radiomic features have not been used clinically in this context yet, many previous studies have investigated DCE-MRI radiomic signatures for the separation of breast cancers of different molecular subtypes, although with mixed results. In our study, a multitude of radiomic features from different feature categories was derived from DWI to capture different aspects of image texture, which might have contributed to the excellent results. The majority of DCE-MRI radiomic studies for the separation of molecular subtypes, however, chose to rely solely on a small, typically COM-based subgroup of radiomic features [14, 16, 44, 45]. Sutton, et al. demonstrated accuracies of up to 89.2 % for the differentiation between molecular subtypes using a combination of pathology data and COM features [14]. Meanwhile, Holli-Helenius achieved AUC values of 0.83–0.88 for the separation of luminal A and B cancers in a small patient collective (n = 27) using COM-features alone [15]. These results are in good agreement with those in our own study (accuracies, 91.5 % and 89.5 %); however, geometric features for shape description performed better than COM features in our study (Fig. 2). In contrast to prior DCE-MRI radiomic studies that show a high level of heterogeneity of scanners and sequence protocols and were only partly successful [4, 12, 13, 16], the acquisition parameters in the present study were strictly homogeneous, which might have had a positive impact on our results. Although DCE-MRI is the most sensitive method for the detection of breast cancer [46, 47], the evaluation of DWI radiomic signatures is of special interest, as they might allow for contrast agent-free tumor biology assessment in times of concern surrounding the safety of gadolinium-containing contrast agents [28‐30].

We acknowledge the limitations of our study. Due to the small number of patients in some subgroups, a further division of our population into a training and validation dataset was not considered appropriate. Instead, we decided to apply a k-nn classifier with leave-one-out cross-validation, an established method that has been used in multiple studies in the field [31, 48, 49]. Our best results in terms of accuracies were achieved for luminal B and HER2-enriched cancers, a finding that may in part be attributed to the small number of patients in these groups (n = 11 and 8, respectively). The mean size of breast cancers included in this study was relatively large, which could explain the high number of triple-negative cancers; at the same time, our results might not be generalizable to smaller lesions in which only few voxel data are available for analysis. Breast cancers were delineated manually, which might introduce a certain level of observer-dependency, and may not be feasible in analyzing large datasets. Furthermore, lesions were segmented two-dimensionally and on the slice with the largest lesion diameter only. Larger prospective studies and multicenter trials with an equal distribution of molecular subtypes are needed to evaluate the robustness of DWI radiomics for the assessment of breast cancer molecular subtypes.

In conclusion, our preliminary data indicate a possible potential that DWI radiomic signatures enable the assessment of molecular breast cancer subtypes and receptor status with high accuracy. DWI radiomic analysis showed higher accuracies when tumor segmentation was performed directly on ADC maps, indicating that solid tumor areas are essential to radiomic analysis. Although DWI radiomic signatures are unlikely to replace formal genetic testing, with additional validation, they might have the potential to provide a noninvasive, contrast-agent-free method to assess tumor biology before and during treatment.