Translational Radiomics: Defining the Strategy Pipeline and Considerations for Application—Part 2: From Clinical Implementation to Enterprise

Abstract

Enterprise imaging has channeled various technological innovations to the field of clinical radiology, ranging from advanced imaging equipment and postacquisition iterative reconstruction tools to image analysis and computer-aided detection tools. More recently, the advancement in the field of quantitative image analysis coupled with machine learning-based data analytics, classification, and integration has ushered in the era of radiomics, a paradigm shift that holds tremendous potential in clinical decision support as well as drug discovery. However, there are important issues to consider to incorporate radiomics into a clinically applicable system and a commercially viable solution. In this two-part series, we offer insights into the development of the translational pipeline for radiomics from methodology to clinical implementation (Part 1) and from that point to enterprise development (Part 2). In Part 2 of this two-part series, we study the components of the strategy pipeline, from clinical implementation to building enterprise solutions.

INTRODUCTION

Radiomics can be defined as the process of uncovering signals buried within images and using such signals to augment the traditional radiologic interpretation and gain insights into the structure, behavior, and therapeutic response profile of a disease. We define translational radiomics as a process of converting the basic radiomic methodologies into evidence-based clinically applicable models that may then undergo the steps of platform standardization, algorithm integration, and applied business intelligence to create a commercialized product for mainstream use.

In this part (Part 2), we identify the steps involved in the process of translating a clinically viable radiomics tool to a successful enterprise solution to include the following components: understanding the most applicable areas for clinical implementation of radiomics, performing outcomes analysis to which radiomic features may be associated, implementing radiomics in a platform that can help guide physician behavior, and identifying ways to continually improve the radiomics algorithms as exponentially increasing data and features are generated in the path toward precision medicine (see Fig. 1).

Fig 1.

The translational radiomics pipeline illustrating all the components thereof.

APPLICATIONS FOR CLINICAL RADIOMICS

Image Interpretation and Direct Image Analysis

Perhaps the most direct application of radiomics from an imaging physician’s standpoint is to improve or facilitate the interpretation of images, as has been investigated in evaluating levels of fibrosis on ultrasound imaging of the liver [1], classifying neurodegenerative disease on MRI [2], and classifying bone tumors on PET-CT [3]. In a sense, this type of application represents an extension of computer-assisted detection and classification and has seen prior direct application in the realms of lung nodule classification or breast cancer detection.

Incorporation of radiomics into other aspects of image analysis stretches the field’s utility into the realm of image-guided therapy. The use of feature analysis for more precise delineation of anatomic boundaries in cardiology or vascular images has the potential to improve the segmentation process for quantitative volumetric analyses themselves or to delineate boundaries or track them for treatment purposes [4–6]. Feature-based tumor boundary delineation has applications in staging as well as radiation oncology treatment planning [7–9].

Precision Medicine—Imaging Phenotypes to Guide Therapy

The realization that diseases and their relationships cannot be understood or treated effectively using traditional disease classifiers has led to the concept of precision medicine. Precision medicine seeks to accurately identify the true nature of disease to apply highly targeted therapeutic agents to achieve an optimal response [10]. The opportunity to bring imaging directly into precision medicine, as both genomics and proteomics and other “omics”-related fields, is one of the most intriguing areas of radiomics’ growth. One of the most aggressively pursued applications of radiomics has been in identifying imaging phenotypes in cancer [11–13]. These phenotypes may or may not have corresponding detectable proteomic or genomic correlates and may be used independently to predict survival, select patients for specific therapies, assess response to systemic or targeted therapies, or otherwise guide therapy [14–22]. By identifying noninvasive means to characterize the whole tumor, rather than based on results of a single needle core biopsy, assessment of the patient’s future risk may be more accurately identified, allowing selection of a more appropriate course of neoadjuvant or adjuvant therapy as well as improved assessment of response to therapy [23–31].

Textural feature analysis has likewise been employed to understand microstructure in nononcologic applications such as cardiovascular disease [32,33]. Feature changes in atherosclerotic plaques [34–37] and abdominal aortic aneurysms [38] have implications for important disease evolution [39–41].

CONNECTING RADIOMICS TO OUTCOMES—HIGHLIGHTS OF THE PROCESS

Radiomics in Clinical Trials

There is a need to incorporate radiomics and radiogenomics into existing and future clinical trials, to codevelop radiomics feature-based criteria for treatment response. For example, because existing treatment response criteria, such as response evaluation criteria in solid tumors, immune-related response criteria, World Health Organization classification, and PET response criteria in solid tumors, do not address the problem of pseudoprogression, integrating radiomics-based response assessment into new immuno-therapy trials should be a priority (Fig. 2).

Fig 2.

Illustration of the factors involved in the outcomes analysis of radiomics. OS = outcomes survival; PFS = progression-free survival.

Incorporation of robust radiomics biomarkers into the design of new clinical trials will increase the rate of convergence onto those feature sets that provide the most information about treatment response. Radiomics can be paired with genomics and circulating tumor deoxyribonucleic acid (liquid biopsy) to develop multidimensional diagnostic-prognostic assessment solutions. A concerted effort between entities such as the Eastern Cooperative Oncology Group and ACR Imaging Network, and between the Quantitative Imaging Biomarkers Alliance and RSNA, is under way to meet these challenges.

Extracting Features to Assess as Predictors of Clinical Outcomes.

The first step toward achieving radiomics analysis of specific volumes that include tumor, other specified lesions, vasculature, or organs is to segment the volume of interest using one of a number of approaches, most of which entail manual, or at best, semi-automated segmentation strategies [42]. A number of such methods have been described by other authors, including region growing methods, level set methods, graph cut methods, active contours (snake) algorithms, and semi-automatic segmentations such as live wires, among others [43–45].

Defining Clinical Outcomes to Which Radiomics Features May Be Compared.

Clinical outcome is often not a single predefined number. Seemingly straightforward concepts, such as overall survival or progression-free survival, are at times difficult to extract from electronic medical records or other informatics sources. Other outcome variables might include more complex concepts such as significant cardiac events, the performance of one or more specified procedures, clinical biomarkers, and so on.

Transforming clinical medical records into outcomes requires extraction of both structured and unstructured data from the electronic health record. Structured information provides specific events or therapies to be placed on clinical timelines relative to the imaging event on which radiomics is being performed. To address the need for standardization of clinical context in the electronic health record in a manner that can be processed in a meaningful way using information technology tools, many clinical and biomedical ontologies have been developed over decades, such as the Systematized Nomenclature of Medicine–Clinical Terms developed by the International Health Terminology Standards Development Organization, which standardizes clinical terminology and physician phrases in multiple languages and is widely used in the exchange of electronic health information [46–50].

The unstructured information in free-text fields (eg, physician’s notes) requires a sophisticated review of clinical notes and other free-text entries to piece together the patient’s medical condition and the timeline of events to achieve quantifiable outcomes, such as progression-free survival in the case of oncology. Basic strategies, such as word tokenization and part-of-speech tagging, can be initially employed; however, higher forms of natural language processing and artificial intelligence (AI)–based methodologies are often required to contextualize events in the medical record [51–58]. Open source natural language processing tools have been developed specifically for application in medical contexts, such as the clinical Text Analysis Knowledge Extraction System and Python Natural Language Toolkit [59–61].

Analyzing Relationships Between Radiomic Feature Sets and Clinical Outcomes.

After understanding the collection of clinical measures that must be made or the genomic or proteomic parameters to which radiomic data are to be correlated, an analysis plan must be constructed. Tests of the stability and reproducibility of features must be incorporated into the study [62]. Radiomics features can number in the thousands, and generally the outcome variables will be drawn from patients numbering an order of magnitude fewer; therefore, care must be taken to avoid the identification of statistically significant findings due to random chance [63]. A number of means of dealing with this issue in the context of this type of discovery research exist [64–68].

Implementing Radiomics in a Platform That Impacts Physician Practice.

Full integration of radiomics in clinical radiology requires a close collaboration between radiologists and computer scientists. Radiologist engagement goes beyond building a pleasing user interface; radiologists will be asked to embrace a shift from a chiefly qualitative use of imaging to a form of imaging assessment in which quantitative assessment is integral. In addition, radiologists will be asked to heavily rely on results obtained through AI methods. Identifying liaisons in the radiology community to educate peers and serve as resources for those wary of AI will both facilitate the translation of radiomics into clinical practice and ensure that the radiology specialty maintains its central role as medicine’s information science specialty.

New technologies, especially if they involve using results from an automated or a semi-automated computational system, can take time before they are universally accepted across medicine. Once results of a radiomics-driven clinical decision support system are clinically validated through outcomes research, it will provide a more robust evidence base to implement radiomics in the practice of precision medicine. Furthermore, in the era of value-based care, evidence of value improvement and cost reduction based on radiomics would further facilitate radiomic’s move into mainstream standard of care.

Crowdsourcing Algorithms

Any data-driven technology requires methodologies to improve based on new information. The roles of deep learning and AI, in general, are integral to radiomics because they enable this type of continuous improvement based on incredibly large sets of new data generated in the clinic each day. The crowdsourcing of algorithms, such as through the popular Kaggle competition website, is a method to obtain insights from a large population of data scientists. Competition participants are often rewarded with recognition and, for top performers, monetary consideration [69], although solutions in the radiomics space will also require a balance between high predictive performance and solution distinctiveness. A large challenge of this crowdsourcing approach is navigating regulatory and institutional requirements in sharing data broadly. Although adding complexity and constraining knowledge discovery, homomorphic encryption may offer a mechanism to address these concerns [70].

Product Development and Product Management

As solutions are developed, they should be tested with feedback from users to allow greater future enhancement, tailored to each health care enterprise system for most optimum use. Closed loop feedback between radiologists and referring physicians will ensure success and future updates to the software tools. Institutions often find it challenging to get buy-in from its users—the medical professionals who deliver health care—because new technology can often mean increased workload, distraction, or risk, even at the promise of better, faster, or cheaper patient care in the future. The design of the software should make it as a seamless part of the work-flow, using principles of user- or human-centered design and iterative software development.

When evaluating the feasibility of radiomics for day-to-day patient care, an institution must consider a strategy to develop and scale a solution, an approach surrounding the adoption of such solution, and the manpower it will take to deploy that solution. Because the field of radiomics is still emerging, a ready solution to put the science into practice may not exist and an institution may likely need to customize proxy solutions, piece together components, or even build some components from the ground up. For any of these cases, the institution will need dedicated product management personnel and an experienced software architect. Scalability considerations will include server and database selection, data and network architecture, as well as the development platform.

Several academic centers and commercial entities have developed radiomics platforms, and ensuring that these platforms produce comparable radiomics feature measurements and that any studies underlying predictive analytics applications follow a set of prescribed guidelines should help to facilitate a coordinated effort to move translational radiomics into the precision medicine mainstream [71–74].

Marketing Radiomics With Comparative Performance Analysis

Driving the adoption of radiomics to the enterprise requires effectively communicating the value proposition to both internal stakeholders and commercial entities. Internal stakeholders, including leaders among clinical practitioners, will benefit from an examination of not only a comparison of the aggregate performance of professional image interpreters versus algorithms, but from a comprehensive analysis of the value of the data added by radiomics and how the performance and value differs according to subject characteristics and the targets of interpretation. CDS tools will need to be available to assist in the organization and interpretation of the volumes of data that radiomics analyses may generate [75].

Mainstream application of any technological advancement in medicine is subject to an evidence-based impact on clinical outcomes and, increasingly, on performance on the value-to-cost model [76]. Radiomics and imaging genomics hold great promise on that front—guiding precision biopsy, predicting and evaluating response to neoadjuvant chemotherapy, predicting response to immunotherapy, and deciphering pseudoprogression. If a radiomics solution is shown to decrease hospital stays or visits, to reduce the number of avoidable imaging studies and interventions, to allow for targeted therapy that avoids trying conventional lines of medication, and, in doing so, to favorably impact the health care financial bottom line, then it will be embraced as an effective as well as efficient state-of-the-art component of health care delivery.

CONCLUSION

There are several key steps in the translation of radiomics techniques to commercially implementable enterprise solutions. Further translation of radiomics to enterprise is still a novel concept, and we have introduced a framework that identifies its unique challenges and attempt to provide a pathway to mainstream application that improves outcomes.

TAKE-HOME POINTS.

• The results of radiomics-based CDS are to be subject to robust outcomes analysis, which then impacts physician behavior.

• The second phase of the translational pipeline involves the steps from clinical application to enterprise solutions, including considerations for algorithm crowdsourcing, product management, and comparative performance analytics.

• The final goal of this process is mainstream application of a clinically viable and commercially successful product that improves the standard and the value of health care.