Image Acquisition

Modern CT, MR imaging, and combined PET/CT units allow for wide variations in acquisition and image reconstruction protocols, and standardization of these protocols across medical imaging centers is typically lacking. This is generally not a problem in the routine identification of radiologic features used in clinical practice. However, when images are analyzed numerically to extract meaningful data, variations in acquisition and image reconstruction parameters can introduce changes that are not due to underlying biologic effects. This has been well recognized in the emerging field of quantitative imaging, in which the intent is to generate medical images with describable limits of bias and variance. In other words, it is not sufficient to report a number or a set of numbers derived from images; instead, we must also be able to provide error bars, as is done with every other credible laboratory measurement.

There have been multiple efforts to advance quantitative imaging, including definition of acquisition and reconstruction standards, over the past 15 years (26,27). The QIN Quantitative Imaging Network is a cooperative network initiated by the NCI National Cancer Institute with the goal of developing quantitative imaging methods that improve the effectiveness of clinical trials of new cancer therapies (28). The QIN Quantitative Imaging Network is a major initiative from the NCI National Cancer Institute and can be regarded as the leading edge of new imaging methods, including radiomics. Also, the Radiological Society of North America (RSNA Radiological Society of North America) and the National Institute for Biomedical Imaging and Bioengineering have sponsored the Quantitative Imaging Biomarkers Alliance (QIBA Quantitative Imaging Biomarkers Alliance), which is a major effort in quantitative imaging (29). The goal of QIBA Quantitative Imaging Biomarkers Alliance is to industrialize quantitative imaging by bringing together the entire spectrum of groups involved in its development and implementation. The main product of QIBA Quantitative Imaging Biomarkers Alliance is a new type of a document termed a profile that provides a consensus on the measurement accuracy of a quantitative imaging biomarker for a specific use and the requirements and procedures needed to achieve this level of measurement accuracy. More than 100 participants were involved in the creation of the initial QIBA Quantitative Imaging Biomarkers Alliance fluorodeoxyglucose PET/CT profile, which was released in 2014 (30,31). The American Association of Physicists in Medicine is providing technical guidelines in quantitative imaging in the form of modality-dependent reports on imager operation and testing. Finally, relevant professional societies, such as the American College of Radiology, RSNA Radiological Society of North America, the Society of Nuclear Medicine and Molecular Imaging, the International Society of Magnetic Resonance in Medicine, and the World Molecular Imaging Society, are increasingly including aspects of the bedrock of quantitative imaging in their guidelines.

Volume of Interest Identification

Identification of tissue volumes of prognostic value is the core of the practice of radiology in oncology. Although at the time of diagnosis cancer can be detected at one tumor site or multiple tumor sites, most patients with cancer metastasis have multiple lesions. In either scenario, we need to identify tumors and suspected tumors as volumes of interest. However, detailed analysis of subvolumes within tumors (the manifestations of tumor heterogeneity) that may have prognostic value generally are not captured in a radiology report because of the spatial and contrast limitations of digital images. While heterogeneity is not included in Response Evaluation Criteria in Solid Tumors, version 1.1 (32), a few texture descriptors have been incorporated in more complex diagnostic imaging reporting and data systems, such as the Breast Imaging Reporting and Data System (BI-RADS Breast Imaging Reporting and Data System) (33), the Prostate Imaging Reporting and Data System (PI-RADS Prostate Imaging Reporting and Data System) (34), and the Lung Imaging Reporting and Data System (Lung-RADS Lung Imaging Reporting and Data System) (35). In the practice of radiomics, so-called subvolumes of interest can be captured and added to the analyses. The basic philosophy, which has its foundation in process engineering, is to capture as much data as possible at the front end and use downstream database mining to identify the features with the highest prognostic value. This is driven by the knowledge that attempting to filter the data at input would be inefficient and would presuppose knowledge regarding the value of the features in classifier models before they were tested.

Recently, the concept of using image data to identify physiologically distinct regions within lesions has been described (36). In this approach, images with different acquisition parameters (eg, contrast material–enhanced T1-weighted MR imaging, diffusion-weighted, and fluid attenuation sequences) can be combined to yield regions with specific combinations of quantitative image data. Notably, when this is performed, the combinations reside in spatially explicit regions of the tumors (Fig 2). We have termed these regions habitats because they represent physiologically distinct volumes, each with a specific combination of blood flow, cell density, necrosis, and edema. Additional radiomic features can be extracted from each of these habitats to obtain highly granular descriptions of cancer lesions. The distribution of these habitats in patients with glioblastoma multiforme, for example, can enable us to discriminate between cancers that progress quickly (<400 days of survival) and those that are more indolent (37). Furthermore, these habitats change after treatment (eg, treatment with radiation and temazolamide), and the pattern of change has been observed to be predictive of response.

Open in a separate window

Figure 2:

Habitats in a patient with glioblastoma multiforme. Habitats were defined by combining unenhanced and contrast-enhanced T1-weighted, 120-msec echo time T2-weighted, and fluid-attenuated inversion recovery (FLAIR) images. Data from each acquisition were sorted into low and high values with automated histogram analyses, yielding a potential for eight different combinations. In practice, only four distinct combinations were observed. They correspond to the red (low T1, high T2 and FLAIR), yellow (low T1 and T2, high FLAIR), blue (High T1 and FLAIR, low T2), and green (high T1, low FLAIR and T2) areas. Notably, while the identities of individual voxels were determined mathematically, they spatially clustered into contiguous regions reflecting different physiologic microenvironments. (Image courtesy of R. A. Gatenby.)

Segmentation

Segmentation is the most critical, challenging, and contentious component of radiomics. It is critical because the subsequent feature data are generated from the segmented volumes. It is challenging because many tumors have indistinct borders. It is contentious because there are ongoing debates over whether to seek ground truth or reproducibility and how much to rely on manual or automatic segmentation. However, a consensus is emerging that truth is elusive and that optimum reproducible segmentation is achievable with computer-aided edge detection followed by manual curation. It is well recognized that interoperator variability of manually contoured tumors is high (38,39). Segmentation of normal structures, such as skeletal elements and organs, can now be achieved with full automation. However, any disease, especially cancer, requires operator input because of inter- and intrasubject morphologic and contrast heterogeneity at the initial examination.

Feature Extraction and Qualification

The heart of radiomics is the extraction of high-dimension feature data to quantitatively describe attributes of volumes of interest. In practice, “semantic” and “agnostic” features are the two types of features extracted in radiomics. (Table 2). Semantic features are those that are commonly used in the radiology lexicon to describe regions of interest, while agnostic features are those that attempt to capture lesion heterogeneity through quantitative descriptors.

Table 2

Examples of Semantic and Agnostic Features of Radiomics

Open in a separate window

Semantic features.—Although semantic features are commonly used by radiologists to describe lesions, in this article we refer to their quantification with computer assistance. With the foreknowledge that semantic features are of prognostic value, early investigations in radiomics developed radiology lexicons, much the same as BI-RADS Breast Imaging Reporting and Data System, PI-RADS Prostate Imaging Reporting and Data System, and Lung-RADS Lung Imaging Reporting and Data System attempt to do. A watershed article in this regard came from Segal et al, who, in an early example of radiogenomics, used a finite series of radiologist-scored quantitative features to predict gene expression patterns in hepatocellular carcinoma (40). This approach continues to have high value, and there is a movement to capture such semantic data with the aid of computers to achieve higher interreader agreement, faster throughput, and lower variance.

Agnostic features.—Agnostic radiomic features on an image are mathematically extracted quantitative descriptors, which are generally not part of the radiologists’ lexicon. These can be divided into first-, second-, or higher-order statistical outputs. First-order statistics describe the distribution of values of individual voxels without concern for spatial relationships. These are generally histogram-based methods and reduce a region of interest to single values for mean, median, maximum, minimum, and uniformity or randomness (entropy) of the intensities on the image, as well as the skewness (asymmetry) and kurtosis (flatness) of the histogram of values. Second-order statistical descriptors generally are described as “texture” features; they describe statistical interrelationships between voxels with similar (or dissimilar) contrast values. Texture analysis of images was first introduced in 1973 by Haralick et al (41). In radiomics, texture analyses can readily provide a measure of intratumoral heterogeneity. In practice, there are dozens of methods and multiple variables that can be used to extract texture features, resulting in hundreds of values—far too many to elaborate on in this article. Readers are referred to some excellent reviews on the subject, with specific reference to intratumoral heterogeneity (42,43). Higher-order statistical methods impose filter grids on the image to extract repetitive or nonrepetitive patterns. These include fractal analyses, wherein patterns are imposed on the image and the number of grid elements containing voxels of a specified value is computed (44); Minkowski functionals, which assess patterns of voxels whose intensity is above a threshold (45); wavelets, which are filter transforms that multiply an image by a matrix of complex linear or radial “waves”; and Laplacian transforms of Gaussian bandpass filters that can extract areas with increasingly coarse texture patterns from the image (46).

There has been a sustained effort to identify, define, and extract more agnostic features. The first such study used 182 texture features in combination with 22 semantic features to describe CT images of lung cancer (24). This was followed by a 442-member feature set that also contained wavelets (47). More recently, this has been expanded to 662 features that also contain Laplace transforms of Gaussian fits (46) and 522 features that include texture and fractal dimension features (48). These features potentially can be extracted from individual habitats, thereby yielding thousands of data elements with which to describe each volume of interest, with many volumes of interest available in each patient.

Thus, it is readily apparent that the number of descriptive image features can approach the complexity of data obtained with gene expression profiling, which commonly yields information on more than 30 000 different sequences. With such large complexity, there is a danger of overfitting analyses, and hence, dimensionality must be reduced by prioritizing the features (49,50). The most systematic approach is to first identify features that may be redundant (ie, those that are highly correlated with one another). Figure 3 is a covariance matrix of 219 features extracted from CT scans in 143 patients with non–small cell lung cancer. Those features that are highly correlated (r² > 0.95) with each other are shown as red off-diagonal elements. Clusters of highly correlated features can be collapsed into one representative feature, usually the one with the largest intersubject variability or highest dynamic range. Figure 3 also provides a conceptual bridge to the other -omics fields, where the data content of the images is indicated by the false-color map. If available, test-retest data are also extremely helpful, as they can help prioritize features on the basis of their reproducibility (51,52). A further level of prioritization, described by Aerts et al (47), is to rank order features within separate categories representing different agnostic and semantic classes of features (eg, size, shape, and first-, second-, and higher-order textures). The classifier models can then be built with the two or three highest-priority features in each class. In the final analysis, the value of feature sets is determined by their contribution to classifier models created through database mining.

Open in a separate window

Figure 3:

Covariance matrix of radiomic features. A total of 219 features were extracted from each non–small cell lung cancer tumor in 235 patients. Across all tumors, each feature was individually compared with all other features by using regression analysis, thereby generating correlation coefficients (R²). Individual features were then clustered and plotted along both axes, and R² is shown as a heat map, with areas of high correlation (R² > 0.95) shown in red. Thus, each of the red squares along the diagonal contains a group of features that are highly correlated with one another and are thus redundant. For data analysis, one feature was chosen to be representative of each of these groups. The representative feature chosen was the one that had the highest natural biologic range (interpatient variability) across the entire patient data set, with the explicit assumption that features that show the highest interpatient variability will be the most informative. (Image courtesy of Y. Balagurunathan.)

Building Databases: Numbers Are King, Quality Is Queen

In radiomics and elsewhere, the power of the predictive classifier model is dependent on having sufficient data. It has been our experience that a reasonable rule of thumb is that 10 samples (patients) are needed for each feature in a model based on binary classifiers. Furthermore, the best models are those that can accommodate additional clinical or genomic covariates, and this increases the need for large high-quality data sets. Radiomics can be performed with as few as 100 patients, although larger data sets provide more power. It is time consuming to capture and curate large high-quality sets from retrospective data. For example, in a recent study, we curated a data set of patients with non–small cell lung cancer adenocarcinoma who had gene expression profiles (46). Within a local database, 285 such patients were readily identified as potential candidates for such a cohort study. The need to validate these via chart and pathology review required 188 hours and resulted in the loss of 50 patients from the study cohort because of missing data or equivocal histopathologic findings. When histopathologic findings were equivocal, a pathologist reviewed the slides; this only added to the curation time. Further validation via picture archiving and communication system review of images captured with standardized acquisition and reconstruction parameters required 94 hours and resulted in the attrition of an additional 92 patients. Segmentation and extraction of features into the database required an additional 145 hours. Thus, the curation of a data set of 143 patients required an initial cohort of 285 patients (approximately 50% attrition) and required 430 hours of processing, or approximately 3 hours of processing per patient. As these patients were not filtered for medical or demographic issues, there was no selection bias. In the future, capturing images and other data prospectively and with higher quality and standards should reduce data attrition and make the process more efficient.

Enabling Diagnosis

In a study of 147 men with biopsy-proven prostate cancer, Wibmer et al (6) showed that Haralick texture analysis has the potential to enable differentiation of cancerous from noncancerous prostate tissue on both T2-weighted MR images and apparent diffusion coefficient (ADC apparent diffusion coefficient) maps derived from diffusion-weighted MR images (Fig 4). In the peripheral zone of the prostate, all five features assessed (entropy, inertia, energy, correlation, and homogeneity) differed significantly between benign and cancerous tissue on both types of images; however, in the transition zone, significant differences were found for all five features on ADC apparent diffusion coefficient maps and for two features (inertia and correlation) on T2-weighted images. In a follow-up study, these features were used to automatically compute Gleason grade and were found to enable discrimination between cancers with a Gleason score of 6 (3+3) and those with a Gleason score of 7 of more with 93% accuracy. Furthermore, these analyses could be used to distinguish between two different forms of Gleason score 7 disease (4+3 vs 3+4) with 92% accuracy (53).

Open in a separate window

Figure 4:

Application of texture analysis to T2-weighted MR images and ADC apparent diffusion coefficient maps of prostate cancer. A lesion in the transition zone is barely discernible on the T2-weighted image (top left) and has higher conspicuity on the ADC apparent diffusion coefficient map (top right). Texture features were computed on a per-voxel basis (using a 5 × 5 × 1 pixel window) from manually segmented regions of interest identifying the normal peripheral zone (outlined in blue) and cancer (outlined in red). From the computed texture features, a machine learning method was applied to distinguish between normal and cancerous structures and to stratify the Gleason patterns. Heat map images show clear differences between healthy tissue and cancer and depict intratumoral heterogeneity that may be useful in assessing tumor aggressiveness and informing fused MR imaging–ultrasonography biopsy.

Tumor Prognosis

Seminal radiogenomic studies were the first to show a relationship between quantitative image features and gene expression patterns in patients with cancer (40,54,55). In the first of these studies, the investigators compared semantic radiologist-defined features extracted from contrast-enhanced CT images in patients with hepatocellular carcinoma to gene expression patterns by using machine learning with a neural network. They found that combinations of 28 imaging traits could be used to reconstruct 78% of the global gene expression profiles, which in turn were linked to cell proliferation, liver synthetic function, and patient prognosis (40). In a subsequent study and with a similar approach, the investigators compared image features extracted from MR images to predict global gene expression patterns in patients with glioblastoma multiforme (54). They found that an “infiltrative” imaging phenotype was associated with significantly shorter survival (54).

In patients with lung cancer, there is incontrovertible evidence for intratumoral heterogeneity on lung CT images (Fig 5). These heterogeneities can be captured with features such as spiculation or entropy gradients. Grove et al found these measures to be strong prognostic indicators in patients with early-stage lung cancer (P < .01) (56). A study by Aerts et al (47) showed that a radiomic signature could be used to predict outcome in completely independent cohorts of patients with lung cancer from two separate institutions. Further, this same signature could be applied to cohorts of patients with head and neck cancer with equivalent prognostic power. Notably, the signature was comprised of the top features from four classes (size, shape, texture, and wavelets) that were prioritized from a database of 442 features by using test-retest reproducibility and intersubject range. This study and others like it demonstrate the potential of radiomics for the identification of a general prognostic imaging phenotype existing in several forms of cancer.

Open in a separate window

Figure 5:

Attenuation gradients of lung CT images. Data are representative of patients with (left) and without (right) recurrence after lobectomy. Although differences in texture were visible on the CT images (top), the color-coded attenuation maps (bottom) more dramatically show intratumoral complexity. Maps were generated by separating the attenuation into quartiles, with hotter colors representing higher attenuation.

It is well known in the radiology community that contrast enhancement at MR imaging is often heterogeneous, with complex patterns. In a landmark article, Rose et al (44) analyzed the pattern of enhancement on dynamic contrast-enhanced MR images in simulations, phantoms, and 23 patients with glioma by using second-order and higher statistical measures to represent enhancement heterogeneity. They convincingly showed that complex measures of texture heterogeneity in transfer coefficient maps from dynamic contrast-enhanced MR images could be used to distinguish high- and low-grade gliomas with much higher statistical power (P < .00005) than could median transfer coefficient maps alone (P = .005). In a more recent study, Gevaert et al extracted a large number of semantic and agnostic features in 55 patients with glioma who had undergone gene expression profiling (57). The feature set was then filtered for reproducibility, yielding 18 features that were assessed in three distinct habitats. Of the agnostic features, most could be correlated with the semantic features; three of 54 were related to survival, and seven were related to gene expression. When gene expression was assessed via pathways, approximately half of the imaging features showed strong correlation to genomics. These analyses show that power for predicting gene expression patterns, outcomes, and staging of gliomas can be significantly increased with radiomics-based approaches.

Recently, Vignati et al performed a thorough prospective radiomic analysis of diffusion- and T2-weighted MR imaging examinations in 49 patients with prostate cancer (58). Agnostic features extracted from T2-weighted images and ADC apparent diffusion coefficient maps were compared with more traditional ADC apparent diffusion coefficient cutoff metrics to test the hypothesis that textures could help differentiate between men with a pathologic Gleason score of 6 and those with a pathologic Gleason score of 7 or higher. This is an important cutoff, as men with a pathologic Gleason score of 6 may be candidates for active surveillance. For standard ADC apparent diffusion coefficient cutoff metrics, the area under the receiver operator characteristic curve ranged from 0.82 to 0.85. When ADC apparent diffusion coefficient and T2 maps were analyzed for heterogeneity, the area under the curve improved to impressive values of 0.92 and 0.96, respectively. Although this study may have been underpowered, it shows the potential value of quantitative analysis of tumor heterogeneity in assessing tumor aggressiveness and informing major clinical decisions, such as whether to treat the cancer at all. Of note, other investigators have also found entropy determined from ADC apparent diffusion coefficient maps to be significantly associated with the pathologic Gleason score, even after controlling for the median ADC apparent diffusion coefficient (6,53).

Treatment Selection

In a seminal study, Kuo et al identified hepatocellular carcinoma imaging phenotypes that correlated with a doxorubicin drug response gene expression program (55). Their results suggested that radiogenomic analyses could be used to guide the selection of therapy for individual tumors. More recently, a study of 58 women who underwent treatment for locally advanced breast cancer suggested that texture analysis of dynamic contrast-enhanced MR imaging could help predict response to neoadjuvant chemotherapy before its initiation (59).

Reproducibility

Radiomics is a young discipline. As with therapies motivated by molecular biology, radiomics offers great potential to accelerate precision medicine. However, it is also possible that radiomics will undergo the same slow progress already experienced with molecular biology–based systemic diagnostic techniques and therapies. That slow progress can be attributed to a number of causes, including technical complexity, poor study design (in particular, mixing hypothesis generation with hypothesis testing) and overfitting of data, lack of standards for validating results, incomplete reporting of results, and unrecognized confounding variables in the databases used, particularly if data are derived retrospectively. Hence, as with any biomarker study, a retrospective radiomics investigation must be validated against a completely independent data set, preferably from another institution. Furthermore, the most rigorous biomarker qualification requires a prospective multicenter trial wherein the biomarker is one of the primary endpoints (62,63).

While standardized tools for genomic profiling (GenomeDx for prostate cancer [GenomeDx Biosciences, San Diego, Calif], Oncotype Dx for breast cancer [Genomic Health, Redwood City, Calif]) have been developed, they are not universally agreed upon or applied across medical centers, hampering efforts to share data and reproduce results. Studies have documented these problems in biomedical research generally and in molecular-targeted drug development specifically. A 2009 analysis of biomedical research reports found that at least 50% of studies were too poor, insufficient, or incomplete to be usable (64). Scientists at Bayer HealthCare (Leverkusen, Germany) reported that they were able to successfully reproduce the published results from only a quarter of 67 seminal studies (65,66). Furthermore, when scientists at Amgen (Thousand Oaks, Calif) tried to replicate 53 landmark studies in the basic science of cancer, they were able to reproduce the original results of just six (67). These issues have become serious enough that editors from more than 30 high-impact-factor biomedical journals have united to impose common standards for statistical testing and to improve access to raw data (68,69). The standards have been adopted by the National Institutes of Health (8,70). Although these standards were generated to address preclinical data, they can be applied across all areas of research and can provide a roadmap for navigating the complex issues associated with acquisition and analysis of high-dimensional data inherent in radiomics. Superb reporting guidelines for clinical studies have been developed by many organizations. An excellent overview is provided by the Equator network, which promotes the quality and transparency of health research (65). Challenges with study design were also identified in the 2012 report Omics from the Institute of Medicine (8). A clear solution to these challenges is to establish benchmarks for the conduct of radiomics studies and for their reporting in the literature.

Big Data

In the era of precision medicine, gigabytes of data are collected for each patient, and radiomics data can provide a significant component of this. The exponential growth in the numbers of patients and the data elements being harvested from each is known colloquially as “big data”. Big data initiatives are aimed at drawing inferences from large data sets that are not derived from carefully controlled experiments. Although correlations among observations can be vast in number and easy to obtain, causality is much harder to assess and establish, partly because it is a vague and poorly specified construct for complex systems. Across big data disciplines there are basic questions: Will access to massive data be a key to understanding the fundamental questions of basic and applied science? Or, does the vast increase in data confound analysis, produce computational bottlenecks, and decrease the ability to draw valid causal inferences? As in radiomics, the field of big data is in its early phases. The aforementioned questions were addressed in a meeting on big data that was sponsored by the National Academy of Sciences (71), and the radiomics field will benefit from this effort.

Data Sharing

The biggest challenge to establishing radiomics-based models as biomarkers to use in decision support is the sharing of image data and metadata across multiple sites. Multisite trials are required to interrogate separate cohorts of patients and to create databases with sufficient size for statistical power. Data sharing is a common challenge in all biomedical research, and it must overcome cultural, administrative, regulatory, and personal issues (72). Notably, communities like the Children’s Oncology Group have established a history and culture of data sharing (73) and therefore are in a prime position to expand their efforts to include radiomic image analysis. Data sharing in radiomics is especially daunting because shared data must include images and sharing must be in compliance with the Health Insurance Portability and Accountability Act, as a substantial amount of personal health information is needed to build models of sufficient complexity. Solutions to this challenge are many and can include: (a) large centralized data repositories, such as The Cancer Imaging Archive and The Cancer Genome Atlas, wherein access can be limited to institutional review board–approved users or data can be stripped of personal health information; (b) federated approaches, wherein each institution maintains their individual data, and query models are sent to extract the relevant metadata; or (c) federated approaches, wherein the institutions are all accessible by an honest broker (ie, a superuser with multisite institutional review board–approved access). No matter which solution is applied, the infrastructure costs can be substantial.

Curation of High-Quality Data by Radiologists

In current general practice, radiologic examinations are qualitatively evaluated, and the generated reports often do not use a standard lexicon, despite a number of efforts to embrace a uniform lexicon, such as RadLex® (74). (If used routinely, annotations with RadLex®-type image features could greatly contribute to mineable databases [75,76].) Furthermore, once images are archived, they are rarely reaccessed. Although massive image repositories exist, they are virtually inaccessible for curation because of the limitations described previously. The only viable solution is to capture data prospectively at the point of care. Hence, we envision a transition from classic radiology to a new paradigm in which the radiologist actively participates in the curation of quantitative image databases. Collection of high-quality image data requires sophisticated content expertise to identity and circumscribe (with computer assistance) and annotate (with a standardized and mineable lexicon) the volumes of interest. To make high-quality data curation a reality, we must first convince the imaging practitioners of its value, and we must streamline the process so it can occur within the limitations of a clinical practice. By playing a crucial role in data curation and analysis of big data, radiologists and physicians alike will be able to make radiomics an important, valuable new dimension of their field.

Health Informatics

To be of maximal value, the various kinds of high-quality data that are obtained during the work-up and monitoring of individual patients must interface with each other. This is well recognized, and most large medical centers are now investing in appropriate electronic medical record systems to make patient data accessible in a mineable form. Currently, radiomic data are typically not incorporated as part of this data stream; however, this is changing with the adoption of structured radiology reporting. The challenge going forward will be to capture radiomic data as part of the structured report.

Data Sharing

As discussed previously, the quality of classifier models is limited by the size of the data sets used to create them. Even if one institution were to capture all of its radiomic data prospectively, it would be years before sufficient power could be generated. Additionally, these data are a moving target, as there are continuous improvements in medical image acquisition. Differences in image acquisition and reconstruction are covariates that must be incorporated in the mining of quantitative image data and therefore increase the amount of data required for model building by an order of magnitude. The solution is for multi-institutional, national, or international consortia to agree to share data either through centralized or distributed (federated) networks. These challenges have been met and solved in the basic sciences of gene expression, sequencing, and protein structure databases. They are beginning to be solved through sharing of oncology metadata (www.oriencancer.org or the Children’s Oncology Group), and some sharing of image files has been enabled by The Cancer Imaging Archive (77).

To date, collaborative efforts to develop quantitative image-driven biomarkers have been remarkable. Radiomics has tremendous potential to further enrich image interpretations and to expand the horizons of imaging toward greater precision and extraction of in vivo biologic information. To fully exploit the potential of radiomics, we need to embrace an interdisciplinary shared vision and make a joint commitment.

The Radiology Reading Room of the Future

The aforementioned scenario will entail a reading room wherein practicing radiologists interact with picture archiving and communication system software to identify, segment, and extract features from regions of interest. If prior studies obtained in the same patient are available, the previous regions of interest will be automatically identified by the reading software. As part of the reading, the extracted size, shape, location, and textural features will be automatically uploaded to a shared database and algorithmically compared with prior images to enable more precise diagnoses. Such capabilities are nearly at hand, as most picture archiving and communication systems have the capability to coregister current images with prior images and perform user-interactive segmentation. For the foreseeable future, the field of radiomics research will be concentrated on improving classifier models to provide the most accurate possible diagnoses and, hence, better patient care and outcomes.