Chapter 6: The Promise of Radiomics

March 4, 2019 | Blog

The Long and Short of It Series

The use of volume, density and other evidence-based volumetric measurements to define lesion characteristics for diagnosis and treatment response monitoring is just the beginning of an evolving revolution in oncology imaging. Lesion segmentation—the process that creates a three-dimensional (3D) region for determining these relatively simple metrics—enables extraction of more sophisticated features that provide a detailed description of a lesion’s morphology.

It is in this space that the relatively new field of study known as radiomics holds great promise. Defined by high throughput extraction and analysis of large amounts of advanced quantitative imaging features, radiomics can help radiologists and oncologists efficiently make the greatest use of volumetric data. Indeed, a growing body of evidence points to the promise of radiomics as a methodology with strong predictive capabilities, complementing the use of more basic features such as diameter, volume and average density.

Notably, one study found that radiomics outperformed the Lung Imaging Reporting and Data System (Lung RADS™)—a widely-used industry standard in lung cancer screening—and volume-only approaches in predicting nodules that would become cancerous. The study found that radiomics predicted patients at highest risk with 93 percent accuracy compared to Lung RADS and the McWilliams Model, which only hit 75 percent and 77 percent respectively.1

Differences in radiomic feature metrics (Entropy, Kurtosis and 3D Laws) can be used to determine whether a lesion is at high risk of becoming cancerous 1

Download the
entire e-book!

If you enjoyed reading this chapter of the Long and Short of It series, you can download the entire e-book by filling in the information below.

The process of radiomics involves first extracting many image features (hundreds or even thousands) from large patient datasets and identifying a subset of features that best predict certain biological properties and clinical outcomes. For example, radiomics researchers interested in lung cancer screening extract features from suspicious lung nodules and combine these data with knowledge of whether each patient did or did not develop lung cancer. They then use advanced statistical and machine learning techniques to determine which individual features best predict malignancy, and ultimately combine these features into predictive models. Such combinations typically include features that describe a nodule’s size, shape, intensity and texture. Texture features are particularly interesting as they are linked to patterns of cellular heterogeneity that may be a signature for malignancy.

Similarly, in another application of radiomics, certain texture features may indicate the presence of sub-populations of cells that confer resistance to therapy.

Via translation of research to clinical practice, radiologists and oncologists can apply prediction models based on radiomics to more accurately diagnose patients and select appropriate treatments.

Radiomics is analogous with and complementary to genomics. Both strive to identify features, whether they be image based or genetic, that provide greater insight into a patient’s clinical situation and needs. More specifically, “imaging genomics” bridges the gap between radiomics and genomics and uses imaging to help determine genomic features.

An example of heterogeneity visualization from an advanced image analysis platform clearly showing different sub-populations of cells.

Dr. Robert Gillies of Moffitt Cancer Center predicts that radiomics is the next frontier in clinical decision making, noting that “radiomics appears to offer a nearly limitless supply of imaging biomarkers that could potentially aid cancer detection, diagnosis, assessment of prognosis, prediction of response to treatment, and monitoring of disease status.” He adds that, “In the foreseeable future, we expect that data gleaned from radiologic examinations throughout the world will be converted into quantitative feature data and that these data will be interfaced with knowledge bases to improve diagnostic accuracy and predictive power for decision support.”2


Radiomics: A Better Road Map for Point-of-Care Decision-Making

Radiomics complements other descriptive reporting techniques by allowing radiologists to extract quantitative features, which can capture subtle differences between and within tumors. These enhanced diagnostic capabilities help confirm whether a lesion or mass is cancer. Once identified as cancer, radiologists can help make predictions about likely prognosis and best treatment options, as well as drill down into the genomic makeup of the cancer, enabling better drug matches and more effective precision medicine.

A flowchart showing the process of extracting radiomics and the use of those radiomics in decision support 2

Recent studies leveraging radiomics and imaging genomics have helped quantify overall tumor spatial complexity and identify subregions that drive disease transformation, progression and drug resistance.3, 4By enhancing prognostic capabilities, these measurements give a better indication of whether a patient will survive.

As well as capturing quantitative features via image analysis, the radiomics process can also include “semantic” descriptions. These are properties that can be described or categorized by a human observer and are traditionally included in radiology reports. Radiomics advances use of semantic descriptions by providing data structures to contain the results and make them readily available for analysis. Examples of semantic descriptors along with possible categories are lesion shape (spherical, lobulated), location (left upper lobe, right lower lobe), borders (smooth, spiculated), attenuation (solid, ground-glass), attached to a vessel (yes or no).

In building prediction models, semantic features can be combined with the “agnostic” features calculated via image analysis. Furthermore, computerized analysis can help in categorizing semantic features, such as automatically determining whether the attenuation of a nodule is below a certain threshold, thus making it a ground glass nodule (GGN). Radiomics-based models can be designed to incorporate semantic descriptors, expressed as categorical or binary variables, along with agnostic features expressed as continuous variables. In short, radiomics helps build a bridge between traditional radiology reporting and a purely quantitative imaging approach.

Examples of agnostic features describing texture include histogram-based metrics, Haralick textures, Laws textures, wavelets, Laplacian transforms, Minkowski functionals and fractal dimensions, helping radiologists drill down into much greater detail than can be revealed by a human observer. For example, radiomics can help accentuate the findings of an attenuation gradient map to reveal which tumors are more heterogeneous compared to those more homogeneous.

Example of an attenuation map that clearly demonstrates intratumoral complexity and how it relates to prognosis 2

Radiomics’ processes begin with image acquisition. Currently modern CT, MR imaging and combined PET/CT units create wide variability in acquisition protocols. Yet, efforts are underway to improve this outlook through the introduction of industry standards.5 Once acquired, a region of interest (ROI) is identified that can encompass either the whole tumor or subregions.

Lesion segmentation is the all-important next step as this process lays the foundation for generating feature data. Traditional manual segmentation – contouring a lesion on multiple image slices – can be very time consuming. Thus, it is not practical in either a radiomics research scenario where a large number lesions need to be segmented, or in a clinical scenario where lesion segmentation is required before applying radiomics-based predictive models. On the other hand, because tumors often do not have well-defined borders, consistent automatic segmentation is elusive. Considerable global scientific effort has gone into development and refinement of segmentation techniques. Fortunately, advanced algorithms that were typically only available in dedicated image analysis software, often “home-grown” within academic institutions and not operating in-line with the clinical workflow are beginning to find their way into clinical solutions. Software enabling sophisticated, efficient segmentation within the clinical workflow are available and will be key in advancing the development and use of radiomics. To be robust, the segmentation tool must include the ability to quickly and intuitively adjust the automatic result as needed.6

Feature extraction comprises the fourth step of the process and is central to achieving the goals of radiomics. Here, high-dimension semantic and agnostic data is obtained and delivered in a report to quantify attributes of a tumor. In a radiomics research scenario, this adds to the body of data used to develop diagnostic, prognostic and predictive models. In a clinical scenario, it is used as input to a selected model that will provide clinical decision support (CDS).


Reading Rooms of the Future

In his paper “Radiomics: Images are More than Pictures, They are Data,” Dr. Gillies points to a reading room of the future where radiologists use software to identify, segment and extract hidden imaging data to improve CDS at the point of care.7 The good news is that tools already exist that automate the process of extracting data related to hundreds of metrics—including size, shape, location, and textural features—and transferring them to a radiomics database. Once uploaded, the information is shared with genomics and medical data within an electronic medical record and then algorithmically compared with prior images to enable more precise diagnoses. Data mining capabilities then refine CDS to specific patient indicators to improve cancer treatment and ultimately, outcomes.

A workflow diagram of the Reading Room of the Future as described by Dr. Robert Gillies in his paper “Radiomics: Images are More than Pictures, They are Data”2is 2

Did you enjoy this chapter of our The Long and Short of It Series?

Click below, in case you missed a chapter.

Chapter 1  |  Chapter 2  |  Chapter 3  |  Chapter 4  |  Chapter 5


  1. S. Hawkins, H. Wang, Y. Ling, A. Garcia, et al. “Predicting Malignant Nodules from Screening CT Scans.” Journal of Thoracic Oncology, December 2016, 11 (12): 2120-2128. Available from:
  2. R. Gillies, P. Kinahan, H. Hricak. “Radiomics: Images are more than Pictures, They are Data,” Radiology, November 18, 2016. Available from:
  3. G. Lee, H. Lee, E Ko, W. Jeong. “Radiomics and imaging genomics in precision medicine,” Precision and Future Medicine, 2017(1): 10-31. Available from: (in references)
  4. G. Lee, et al. 2017.
  5. R. Gillies, et al. November 28, 2016.
  6. E. Rios Velazquez, HJ Aerts, Y. Gu, et al. “A semiautomatic CT-based ensemble segmentation of lung tumors: comparison with oncologists’ delineations and with the surgical specimen.” Radiotherapy Oncology, 2012;105(2):167–173.
  7. R. Gillies, et al. November 28, 2016.