Images Are Data: Challenges and Opportunities in the Clinical Translation of Radiomics

Abstract

Radiomics provides an opportunity to uncover image-based biomarkers through the conversion and analysis of standard-of-care medical images into high-dimensional mineable data. In the last decade, thousands of studies have been published on different clinical applications, novel analysis algorithms, and the stability and reproducibility of radiomics. Despite this, interstudy comparisons are challenging because there is not a generally accepted analytic and reporting standard. The ability to compare and combine results from multiple studies using interoperative platforms is an essential component on the path toward clinical application. The NCI supported study from van Griethuysen and colleagues published in Cancer Research in 2017 proposed PyRadiomics: an open-source radiomics quantification platform for standardized image processing. Since released, it has become a frequently utilized analytic tool in the radiomics literature and has accelerated the capability of combining data from different studies. The subsequent challenge will be the design of multicenter trials with a fixed and immutable version of software, which is currently open-source, readily modified and freely distributed. Generally, this is accomplished with a commercial partner to navigate the regulatory processes. See related article by van Griethuysen and colleagues, Cancer Res 2017;77:e104-7.

Figure 1.. Radiomic publications from 2012 to present by:

(A) cancer site, (B, C) imaging modality, (D) hand-crafted radiomics vs. deep learning, (E) study types, and (F, G) sample size by imaging modality.

Figure 2.. Image-based analytical pipelines and qualification for radiomic biomarkers.

Conventional hand-crafted radiomic features are calculated from user-defined volumes of interest (VOI) through a series of steps black and green boxes) that include volume delineation, feature extraction/calculation, feature selection/reduction, and then task-oriented predictive model development and validation (right end of the blue box). Deep learning (blue box) methods that do not depend on volume delineation of an VOI and radiomic features are not extracted. Rather, a patch (e.g., bounding box), an entire image (i.e., slice), or an entire volumetric image series can be an input into a deep learning network for the task-oriented predictive model development and validation.