The goal of this work was to evaluate whether directional-gradient based radiogenomic descriptors on DCE-MRI were correlated with molecular subtypes in HER2+ breast cancers. Radiogenomics refers to the process of attempting to identify associations between radiomic and genomic features of tumors in order to describe and predict tumor biology, molecular subtype or mutational status. HER2+ breast cancer is biologically heterogeneous, and over 50% HER2+ breast cancers either progress or become resistant to trastuzumab-containing therapy. Given that HER2 amplicon drives cancer progression within luminal and basal cellular lineages, PAM50-based gene expression subtyping¹ of HER2+ breast cancer results in three major subgroups - HER-Enriched, HER2-Luminal and HER2-Basal. We recently demonstrated, that HER2-Enriched subtype consistently exhibits the highest rate of pathologic complete response to preoperative anti-HER2 therapy². Use of textural kinetics to capture hormone receptor status has been shown in³. We hypothesize that the differences in underlying biology at a cellular/molecular scale is also reflected at the imaging scale. In this work, we present the initial findings of using a directional gradient-based, computer-extracted radiogenomic descriptor, Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe), to characterize HER2+ breast cancer and provide insight into the underlying structural and biological heterogeneity. The ability to non-invasively identify HER2+ subtypes based on sub-visual radiogenomic signatures may allow for improved understanding and better clinical management of breast cancers.CoLlAGe is a radiomic feature that enables quantitative measurements of the degree of order/disorder of pixel-wise localized image gradient orientations. Differential expression of CoLlAGe in pathologies of varying degrees of aggressiveness has been shown in the context of both brain tumors and breast cancer⁴. Our dataset consists of 25 DCE-MRI breast cancer (from the completed BrUOG-211B trial) cases with subtypes identified using the PAM50 gene expression signature. Hierarchical clustering using PAM50 genes was used to identify sub-groups corresponding to ER/PR immunohistochemistry and other proliferation genes. The resulting distribution of subtypes included 10 HER2-Enriched, 12 HER-Luminal and 3 HER2-Basal cases. Using a 1.5/3.0 T magnet, STIR axial and T1w fat saturation axial scans, before and after intravenous Gd-contrast administration, were obtained with an 8 or 16 channel dedicated breast coil. The lesions were manually delineated on the peak enhancement phase followed by pixel-wise computation of gradient orientations on the annotated regions of interest. Local dominant orientations were then computed via principal component analysis and entropy features were extracted on a per-pixel basis from the co-occurrence matrix of the dominant orientations. We also extracted DCE-MRI pharmacokinetic (PK) features (K^trans: volume transfer coefficient, K^ep: flux rate constant, V^e: extracellular volume ratio) using Tofts model⁵, as a comparative strategy. Figure 1 shows the differential expression of CoLlAGe entropies across the subtypes. Using first order CoLlAGe entropy statistics, groups were identified in a hierarchical unsupervised cluster setting (Figure 2(a)). Similarly using PK parameters and intensity statistics, groups were identified in Figure 2(b). Mean of CoLlAGe entropy was able to distinguish HER2-Enriched from basal+luminal while kurtosis and skewness were able to distinguish HER2-basal from enriched+luminal. The three distinct clusters in Figure 2(a) were the three different subtypes, as identified by the PAM50 assay, with enriched+basal clustering accuracy of 70%. Interestingly, the corresponding clustering accuracy with PK parameters and signal intensity was only 54% with no distinct clusters correspoding to HER2+ subtypes (Figure 2(b)). The ground truth comprises the labels obtained from the PAM50 gene expression signature. The results suggest that CoLlAGe can be a potential surrogate biomarker in breast cancer imaging because 1) on baseline imaging (pre-biopsy) it can stratify HER2+ cases into distinct subtypes based on cellular lineage hormone receptor status and 2) it provides greater understanding of the associated biological heterogeneity within the tumor microenvironment across subtly different subtypes. Radiologic phenotypes are a reflection of cellular/molecular phenotypes. While PK parameters quantify intra-tumoral permeability changes, CoLlAGe attempts to capture degree of order in pixel-level gradient orientations within local neighborhoods of the tumor habitat, thus enabling quantification of subtle micro-textural changes and tumor heterogeneity that may not be captured by MRI signal intensity or PK parameters. Through this radiogenomic approach, we attempt to establish associations between radiomic and genomic features of HER2+ breast cancer in order to describe and predict underlying tumor characteristics. Subtype identification has a high prognosis impact as it can provide insights into therapy response, and can potentially help in improved clinical management of breast cancers. Our radiogenomic approach combined with routine assays can potentially result in novel biomarkers of therapy response, thereby enabling more effective guided therapy.