“Radiomics” refers to the extraction and analysis of large amounts of advanced quantitative imaging features from medical images obtained with computed tomography (CT), positron emission tomography (PET) or magnetic resonance imaging (MRI) using high throughput methods.¹ The suffix -omics originates in molecular biology to describe the detailed characterization of biologic molecules such as DNA (genomics), RNA (transcriptomics), proteins (proteomics), and metabolites (metabolomics). Hence the origin of the ‘new buzzword’ – radi-omics as it applies to the process of extraction and analysis of the hundreds quantitative imaging variables in radi-ographic images.

Radiomic data are in a format that is amicable for building descriptive and predictive models relating image features to a variety of patient characteristics. Resultant models may include imaging, molecular, and clinical data, and can provide valuable diagnostic, prognostic or predictive information. The process of conversion of digital medical images into mineable high-dimensional data is motivated by the concept that biomedical images contain information that reflects underlying pathophysiology and that these relationships can be revealed via quantitative image analysis. And indeed, quantitative image features based on intensity, shape, size or volume, and texture offer information on tumor phenotype that is distinct from that provided by clinical reports, laboratory test results, and genomic or proteomic assays. These features, in conjunction with the other information, can be correlated with clinical outcome data and used for evidence-based clinical decision support (Figure 1).²

Quantification and characterization of radiomics features have been found to reflect tumor molecular characteristics (radiogenomics) and, hence, heterogeneity in solid tumors.³ Specific quantitative MRI parameters have been significantly related to molecular subgroups in glioblastoma multiforme (GBM).^4-6 Likewise, using the cancer genome atlas (TCGA) renal cell carcinoma dataset, Shinagare et al. correlated margin and growth pattern imaging parameters with BAP1 and MUC4 mutation, respectively.⁷ In another study of hepatocellular carcinoma, dynamic CT imaging traits were shown to be associated with global gene expression profiles and only 28 imaging features were required to reconstruct all gene expression clusters observed.⁸ Most recently, Aerts et al. defined 440 radiomic features and applied these to 1019 lung and head and neck cancers to show that radiomic data had both prognostic power, as well as strong association with gene expression patterns.⁹

The syllabus is structured as follows: First, MRI radiomic features and extraction are described; second, examples of applications of radiomics in GBM and prostate cancer are reviewed and lastly the importance of incorporating radiomics features in clinical databases is discussed.

The two main steps in the radiomic process are: (i) segmenting the volumes of interest; and (ii) extracting and qualifying descriptive features from the volume.

Segmentation: Segmentation is the most critical, challenging, and contentious component of radiomics.² The first step is to determine what is the volume of interest – the enhancing lesion of the GBM or the edema; the entire organ, as in the case of the prostate or a sub volumes of the tumor. This process is crucial because the subsequent feature data are generated from the segmented volumes. The delineation of the volume of interest is challenging because it requires an operator (radiologist) input to contour the volume introducing subjectivity in the process. For example, the ‘spiculation’ of the tumor is highly operator dependent. It is well recognized that interoperator variability of manually contoured tumors is high. Major efforts are devoted to developing automatic segmentation.^10,11 However, a consensus is emerging that the truth is elusive and that optimum reproducible segmentation is achievable with computer-aided edge detection followed by manual curation.

Feature Extraction: Once the volume of interest is outlined, several types of image features are extracted to describe the volume shape and structure.^1,12 Such features are illustrated broadly in Table 1. (C1) Volume descriptors can be calculated for multiple structures; (C2) Quantitative features describing geometric shape will be extracted from the 3D surface of the rendered volume; (C3) Tumor location (for example, for prostate on a standard 6-volume template will be recorded); (C4) Intensity histogram based features reduce the 3D data into a single histogram and common statistics for each region/modality may be calculated (e.g. mean, median, min, max, range, skewness, and kurtosis); (C5) Co-occurrence matrix features are widely used for texture classification. The joint conditional probability density function P(i.j;a,d) is the basis of co-occurrence matrix: the elements (i,j) represent the number of times that intensity levels i and j occur in two voxels separated by the distance d in direction a. Subsequently, from this conditional probability density function, features can be extracted, e.g., describing autocorrelation, contrast, cluster prominence, cluster shade, cluster tendency, dissimilarity, energy, homogeneity, maximum probability, sum of squares, sum average, sum variance, sum entropy or difference entropy, etc.

GBM: GBM is the most aggressive and rapidly fatal primary brain tumor resulting in a poor prognosis and a 15 month survival rate on average.¹³ In a pioneering study, Diehn et al.⁶ associated imaging and gene expression modules in GBM. Ten binary imaging traits were scored for 22 MRI studies of GBM patients by a neuro-radiologist: contrast enhancement, necrosis, mass effect, pattern of T2 edema (infiltrative/ edematous), cortical involvement, Subventrical Zone Involvement (SVZ), Contrast to Necrosis ratio, Contrast to T2 ratio, degree of T2 edema, and T2 heterogeneity. These features are also known in radiomics as ‘semantic’ (as opposed to ‘agnostic’, that are the automatically extracted quantitative descriptors, defined in Table 1). Semantic features are those that are commonly used in the radiology to describe regions of interest. The ten imaging features in Diehn et al.⁶ were related to gene-expression profiles assessed by 2188 cDNA clones microarray analysis. As a result, tumor contrast enhancement and mass effect predicted activation of specific hypoxia and proliferation gene-expression programs, respectively. Overexpression of EGFR, a receptor tyrosine kinase and potential therapeutic target, was also directly inferred by neuroimaging and was validated in an independent set of tumors by immunohistochemistry. The “infiltrative” imaging phenotype was associated with survival-associated gene signature and in a larger group of patients this imaging phenotype predicted patient outcome (Figure 2).

Predictive models for survival by correlating qualitative imaging phenotypes were built by correlating quantitative imaging features with gene expression data alone14 or with addition of microRNA data.15 Recently, Itakura et al.¹⁶ sought to identify subtypes of GBM, differentiated solely by quantitative MR feature, that could be used for better management of GBM patients. Quantitative image features capturing the shape, texture, and edge sharpness of each lesion were extracted from MR images of 121 single-institution patients with a de novo, solitary, unilateral GBM. Three distinct phenotypic "clusters" emerged: pre-multifocal, spherical, and rim-enhancing (Figure 3). These clusters were validated in an independent cohort consisting of 144 multi-institution patients with similar tumor characteristics from The Cancer Genome Atlas (TCGA). Each cluster mapped to a unique set of molecular signaling pathways using pathway activity estimates derived from the analysis of TCGA tumor copy number and gene expression data with the PARADIGM (Pathway Recognition Algorithm Using Data Integration on Genomic Models) algorithm. Distinct pathways, such as c-Kit and FOXA, were enriched in each cluster, indicating differential molecular activities as determined by the image features. Each cluster also demonstrated differential probabilities of survival, indicating prognostic importance. This imaging method offers a noninvasive approach to stratify GBM patients and also provides unique sets of molecular signatures to inform targeted therapy and personalized treatment of GBM.

Prostate Cancer: Prostate cancer is the most prevalent male malignancy in the US, with more than 1 in 6 men expected to be diagnosed with the disease in their lifetime. Treatment recommendations are currently based on risk stratification using prostate specific antigen (PSA), Gleason score and T-category, which typically categorize men as having low, intermediate, or high risk disease.¹⁷ The overtreatment of men with prostate cancer is a well-recognized problem and active surveillance has rapidly become a standard recommendation for many men with low risk disease.¹⁸ Recently, an analysis of 17,943 patients considered candidates for active surveillance and who underwent radical prostatectomy (RP) in the US between 2010-2011 showed that upgrading, upstaging, or nodal metastases occurred in 45% of men and that the deferral of radical prostatectomy for more than 12 months was associated with an 1.7-fold increased risk of non-organ confined disease after surgery.¹⁹ Deferral of radical prostatectomy has also been associated with a higher rate of biochemical progression.²⁰ New methods are needed to improve risk stratification and optimize management.²¹

Recently, Wibmer et al.²² showed that several texture features (also known as Haralick features) appear useful for prostate cancer detection and Gleason Score assessment. Tumor Energy and Entropy on Apparent Coefficient Diffusion (ADC) maps, derived from Diffusion Weighted Imaging (DWI) correlate with Gleason score. In a follow up study, these features were used to automatically compute Gleason Grade and were found to enable discrimination between cancers with GS 6 and >7 with 93% accuracy.²³

Stoyanova et al.²⁴ investigated whether quantitative imaging features were associated with gene expression pertinent to prostate cancer adverse outcome and other mechanisms. While genomic classifiers have improved patient risk classification over standard clinicopathological variables, the multifocality and heterogeneity of prostate cancer can confound pre-treatment assessment. Using quantitative multiparametric (mp)MRI parameters spatially distinct areas (‘habitats’) based on differential perfusion and diffusion characteristics were identified. Global gene expression profiles were generated from 17 mpMRI directed prostate biopsies using an Affimetrix platform. Strong associations between 49 radiomic features and gene expression of 22-gene Decipher panel were found (Figure 4). Decipher® (GenomeDx, San Diego, California)^28-31 is established commercial gene-signature associated with prostate cancer prognosis. Radiomic features segregated the genes into those that were up or down-expressed in more aggressive cancers. Two-way hierarchical clustering of Pearson’s correlation distances between the genomic and radiomic features also revealed strong correlations with 212 genes selected by expression criteria; notably, this unbiased gene-set was markedly enriched with prostate cancer related genes. Gene ontology (GO) analysis identified specific radiomic features from tumor and normal appearing tissue that were associated with distinct biological processes in biopsied tumor tissue related to immune/inflammatory response, metabolism, cell and biological adhesion.

Imaging is crucial for clinical decision making, and yet it is poorly integrated with clinical and molecular data in contemporary clinical databases. Hundreds of diagnostic and treatment variables are recorded routinely in clinical evaluation while the results from imaging exams are reduced to handful of categorical values. Difficulties include that the information in the images is neither explicit nor in standard computer-accessible formats. The problem is further complicated by multiple imaging modalities, creating the need to segment, map, co-register in space and time, and fuse. Each action is a source of errors in accuracy and reproducibility.

At the University of Miami (UM) we partnered with MIM (MIM, Cleveland, OH), a commercial platform for image processing to develop a systematic procedure for semi-automated processing of mpMRI into explicit, standard computer accessible format, and integrate the quantitative imaging features in existing clinical RedCap database. We are refining quantitative imaging methods that enhance tumor characterization and biopsy positioning through the application of spatially explicit quantitative image analysis that recognizes tumor heterogeneity and defines regionally distinct habitats.²⁵

The nascent field of radiomics has the potential to enable the addition and analysis of computer accessible imaging features to diagnostic, pathologic, therapeutic, and follow-up data in clinical trials that may lead to better stratification, treatment, and prognosis of patients, and lead to better insights into imaging features that are non-invasive biomarkers of underlying molecular pathways, and overall, better future prospective clinical trials. Although radiomics can be applied to a large number of diseases, it is developing at large pace in oncology because of the support of the National Cancer Institute (NCI) Quantitative Imaging Network (QIN) and other initiatives from the NCI Cancer Imaging Program.