Current status of Radiomics for cancer management: Challenges versus opportunities for clinical practice

Radiomics, the high‐throughput extraction and analysis of features from medical images, is a promising field for characterizing tumor phenotype and normal tissue injury post‐radiotherapy. Radiomics provides unique opportunities to identify predictive and prognostic imaging biomarkers in noninvasive imaging assays providing so‐ called digital biopsies that can be acquired throughout the whole course of cancer treatment. Radiomics have been proved to be associated with underlying gene expression and therapy response, which is an area currently referred to as radiogenomics. Multimodality imaging biomarkers extracted from positron emission tomography (PET), computed tomography (CT), magnetic resonance imaging (MRI), and images in other medical modalities have been shown to have discriminative power for cancer treatment outcome prognosis and prediction. For example, F‐fluoro‐2‐deoxy‐D‐glucose (FDG)‐PET images are the standard of care in tumor quantification of head and neck radiation therapy (RT) and will likely remain so for the foreseeable future. Metabolic tumor volume, defined as the volume of tumor tissues with increase and heterogeneous FDG uptakes, is an important prognostic factor in many malignancies. The radiomics features can complement known first order imaging biomarkers and provide further insights beyond those revealed to naked eyes from medical images. During the past years, there has been tremendous growth in the radiomics field leading to improved performances in cancer diagnosis, cancer staging, tumor classification, treatment outcome prediction, patient survival, and other clinical practice, compared to other simple clinical biomarkers such as tumor staging, tumor size, human papillomavirus (HPV) status, etc. Clinical applications of radiomics have been widely investigated as well. Radiomics yield great promise to support clinical practice and achieve many promising results. There are many publications and special issues dedicated to the usage of radiomics to support clinical applications in combination with recent spread of advanced machine learning methods. Yet questions remain if the development of radiomics makes it ready for prospective clinical use. Herein, we brought in two medical physics experts both of whom have extensive knowledge in clinical practice and radiomics research. Dr. Hua Li is taking the proposition that “Radiomics poses more challenges than opportunities for clinical practice in cancer management,” whereas Dr. Issam El Naqa argues against it. Dr. Hua Li is currently a research associate professor in the Department of Bioengineering at University of Illinois at Urbana‐ Champaign and a clinical medical physicist at Carle Cancer Center, Carle Foundation Hospital, Urbana, IL. Before joining UIUC and Carle, she was an associate professor in the Department of Radiation Oncology at Washington University in Saint Louis. Dr. Li is certified in Therapeutic, Diagnostic, and Nuclear medical physics by the American Board of Radiology. She has conducted active research in developing advanced machine learning, pattern recognition, and image analysis techniques for applications in radiation therapy and diagnostic imaging. Her current research projects include radiomics‐ based prognostic model of cervical cancer habitats, multimodal biomarkers for personalized oropharyngeal cancer treatment, and task‐based image quality assessment and optimization in radiation therapy. Her research projects are funded by the National Institute of Health (NIH). Dr. EI Naqa worked as a Professor and associate member in Applied Physics and the Michigan institute of data science. He recently accepted the position of founding chair of the department of Machine Learning at Moffitt Cancer Center, Tampa, Fl, where will officially start later this summer. He is a certified Medical Physicist by the American Board of Radiology. He is a recognized authority in the fields of machine learning, data analytics, and oncology outcomes modeling and has published extensively in these areas with more than 180+ peer‐reviewed journal publications and four edited textbooks. He has been a senior member and fellow of several academic and professional societies. His research has been funded by several federal and private grants in Canada and the USA and served on national and international study sections. He acts as a peer‐reviewer and editorial board member for several leading international journals in his areas of expertise.


2.A | Hua Li, PhD
The usage of radiomics for reliably and efficiently supporting clinical decision-making in cancer therapy remains largely immature with many impediments. An effective yet stable methodology model able to select and learn from radiomics (or deep learn from the related images) to support clinical practice of cancer diagnostics, treatment prognosis, and prediction is still desired. 5,6 In this section, we highlight the main challenges with the hope to guide the design and development of efficient and robust methodology to address these challenges and expedite the clinical applications of radiomics.

2.A.1 | Limitations of medical imaging systems
Imaging systems are imperfect and generally acquire indirect measurements of an object property that is affected by multiple sources of noise. Tomographic images reconstructed from such measurements are additionally influenced by the choice of reconstruction method and may contain artifacts. The spatial resolution of imaging systems is limited by the imaging physics and the instrument response. A fundamental question is whether a given imaging system is capable of producing images that contain information that is potentially useful for making predictions related to treatment outcomes or other tasks. For example, high heterogeneity of radiomics exists among images of the same patient but acquired from different imaging systems. 7 Moreover, the lack of harmonization and standardization for radiomics due to the limitations of imaging systems is an intrinsic issue that needs to be aware of.

2.A.3 | Radiomic features yield high heterogeneity
Radiomic features, extracted from multimodality images and individual patients, yield high heterogeneity due to varied imaging system principles, numerous imaging protocols and parameters, and intrinsic differences among images from different modalities. The high heterogeneity of radiomics brings additional challenges to determine stable and informative features and properly fuse them to support clinical applications of radiomics.

2.A.4 | Radiomic features yield high uncertainty
Uncertainties and variations in delineated tumor shapes and sizes cause high uncertainty of extracted radiomics from images in different imaging modalities and exhibited in individual patients, mainly due to variations among manual, semi-automatic, or automatic tumor delineation methods.   Therefore, medical image-specific network design, training, validation, and testing are required for the safe use of DLRs to guide clinical practice.

2.A.5 | Uncertainty in clinical outcomes
In summary, imaging system limitations, data redundancy, heterogeneity, uncertainty, and imbalance of the large number of candidate radiomics features, and existing small training sets are the challenges that need to be aware of. A critical barrier hampering the widespread and stable use of imaging biomarkers in clinical practice is the lack of robust tools for identifying prognostic biomarkers from high-dimensional features that work across patient population. The employment of traditional machine learning and advanced deep learning methods in medical imaging fields also requires some special considerations.

2.B | Issam El Naqa, PhD
Despite the aforementioned challenges, there is light at the end of the tunnel. Moreover, delaying the adoption of radiomics until all these challenges are resolved is impractical and will miss the current existing opportunities for employing radiomics to support clinical decision-making. The mere presence of noninvasive nature of medical images and possibility of high spatial and temporal resolution provide major benefits over using simplistic metrics that would overlook the wealth of useful information on tumor's shape, growth/ shrinking over time, heterogeneity that radiomics can provide as discussed below.

2.B.1 | New standards and metrology for radiomics are evolving
Many organizations are leading the way to standards for quantitative imaging and its utilization as biomarkers including the RSNA and AAPM, among others. These initiatives and others have led to useful recommendations for repeatability and reproducibility such as the Quantitative Imaging Biomarkers Alliance (QIBA) in MRI and other modalities, for instance. 9 In addition, consortium such as The Image Biomarker Standardization Initiative (IBSI) specifically designated for standardizing radiomics for high-throughput image-based phenotyping. 10

2.B.2 | Multimodality imaging is underutilized
The use of imaging in RT is progressing at a rapid pace. The simplistic use of this existing wealth of imaging modalities is a typical case of information waste that is missing tremendous opportunities that radiomics can bring to bear beyond simple intensity or volumetric metrics that currently dominate the radiological lexicon. This is cur- 2.B.4 | Advances in data science are benefiting radiomics Though radiomics have its own issues like redundancy, uncertainty, and instability like any other type of 'omics (genomics, transcriptomics, and proteomics), which did not limit their utilization and should not limit radiomics as well. Indeed, dealing with large number of variables is major challenge in statistical modeling and machine learning. However, advances in machine learning and deep learning specifically opened the door for complementing feature-based methods with featureless (machine learnt) methods that are rapidly advancing and would make the future of radiomics even brighter.
In short, radiomics has its own share of challenges but these should not be a hindrance to its cautious use, given the facts that radiomics benefits far outweigh its risks and the missed opportunities that it can offer.

3.A | Hua Li, PhD and Issam El Naqa, PhD
In spite of those abovementioned challenges, there are promising opportunities of continuously employing radiomics to support clinical decision-making considering the unique image characteristics such as EDITORIAL