Introduction

Gene expression profiling has revealed four main intrinsic molecular subtypes of breast cancer that show differences in phenotype, prognosis, treatment response and outcome: luminal A, luminal B, human epidermal growth factor receptor 2 (HER2)-enriched and triple negative [1]. Currently, the assessment of molecular subtypes to guide treatment selection is mostly based on immunohistochemical surrogates from invasive tissue sampling. However, this approach is limited, as biopsy can only capture a small part of a potentially heterogeneous lesion, and tumor biology is subject to change over time. The underlying idea of radiomics is that biomedical images carry information that is reflective of processes occurring at the genetic and molecular level. Features that can be assessed by advanced image processing and analysis, and which are beyond the perception of the human eye, may reflect underlying phenotypic and genotypic tissue characteristics, and might enable noninvasive spatio-longitudinal monitoring of the entire tumor. Previous radiomics studies in breast cancer have primarily focused on features derived from dynamic contrast-enhanced MRI (DCE-MRI). With the recent controversy about the safety of gadolinium-containing contrast agents and the recommendation to use these only when essential information cannot be obtained with unenhanced scans, the investigation of diffusion-weighted imaging (DWI) coupled with radiomic analysis is of special interest. Although DWI has been implemented in breast imaging [2,3], data is scarce about DWI radiomic signatures and limited to the differentiation between benign and malignant breast tumors [4,5]. The aim of this study was to expand on this knowledge and to evaluate the diagnostic performance of DWI radiomic signatures, using different approaches of tumor segmentation, for the assessment of breast cancer receptor status and molecular subtypes.

Methods

Ninety-one patients with treatment-naïve, biopsy-proven breast cancer (luminal A, n=49; luminal B, n=8; HER2-enriched, n=11; triple negative, n=23), who underwent state-of-the-art multiparametric 3T MRI were included in this IRB-approved HIPAA-compliant retrospective study. Radiomic analysis was performed exclusively using ADC maps. Two approaches of manual, two-dimensional tumor segmentation were compared: (a) all lesions were segmented on high b-value DWI and the respective regions of interest (ROIs) were copied to the corresponding ADC maps; and (b) a subgroup of tumors (n=79) was directly segmented on the ADC maps. Radiomic analysis included the following features: first-order histogram (HIS), co-occurrence matrix (COM), run-length matrix (RLM), absolute gradient (GRA), autoregressive model (ARM), discrete Haar wavelet transform (WAV), and lesion geometry (GEO). Fisher, probability of error and average correlation (POE+ACC), and mutual information (MI) coefficients were used for feature selection. Linear discriminant analysis for dimensionality reduction was followed by k-nearest neighbor classification with leave-one-out cross-validation for pairwise separation between molecular subtypes/receptor status. Histopathology served as the standard of reference.

Results

Tumor segmentation on DWI yielded the following accuracies >80%: luminal B vs. HER2-enriched, 94.7% (MI/based on COM features); luminal B vs. others, 92.3% (Fisher/COM, HIS); HER2-enriched vs. others, 90.1% (Fisher/RLM, COM); luminal A vs. luminal B, 89.5% (MI/COM); HER2-enriched vs. luminal A, 83.3% (Fisher/COM, RLM). Better accuracies were achieved with segmentation directly on the ADC map: luminal A vs. luminal B, 91.5% (POE/COM, WAV); luminal B vs. HER2-enriched, 100% (Fisher/COM, WAV); luminal B vs. triple negative, 89.3% (POE/COM); luminal B vs. others, 91.1% (Fisher/WAV, ARM, COM); HER2-enriched vs. luminal A, 80.4% (Fisher/COM); HER2-enriched vs. triple negative, 81.3% (POE/COM); hormone receptor positive vs. HER2-enriched, 84.2% (MI/COM). The twelve lesions which could be not segmented directly on the ADC map included smaller cancers (mean size, 2.7 vs. 3.7 cm), NMEs and/or very dense breasts.

Discussion

To our knowledge, this is the first study to examine the utility of DWI radiomic signatures for the differentiation between breast cancers of different molecular subtype/receptor status. Our preliminary study demonstrates that DWI radiomic signatures enable the assessment of molecular subtypes and receptor status with high accuracy. Classification accuracies were generally higher when tumor segmentations were performed directly on the ADC maps. As ADC maps are generated using DWI signals at 2 different b-values and cancel out T2-shine-through, better tumor delineation is facilitated and therefore seem to be the preferred approach for radiomics analysis. Larger prospective studies are warranted to validate these initial promising findings.

Conclusion

DWI radiomic signatures enable the assessment of breast cancer receptor status/molecular subtypes with high accuracy. Higher accuracies are achieved when tumor segmentations can be segmeted directly on ADC maps, and may therefore be the preferred approach for radiomic analyses. Initial results indicate that DWI radiomic signatures may have the potential to provide information derived from the entire tumor and may be used to monitor spatio-longitudinal tumor biology changes during treatment.

Acknowledgements

No acknowledgement found.