Feature Extraction & Radiomics
1Medical Imaging, Sunnybrook Health Sciences Center, 2University of Toronto, Toronto, ON, Canada
Synopsis
Radiomics is defined as: " “conversion of digital medical images into mineable high-dimensional data… motivated by the concept that biomedical images contain information that reflects underlying pathophysiology and that these relationships can be revealed via quantitative image analyses”. Radiomic features are comprised of imaging biomarkers (IB)Some key questions must be answered at an early stage: “Does the IB fulfill an unmet clinical need?”; “Does data exist to evaluate the IB and if not can it be obtained?”. At an early stage, technical validation including assessment of precision through repeatability and reproducibility must be determined. Furthermore, biologic and clinical validation must also be performed. Cost effectiveness must also be considered. The paradigm and consideraiton in radiomics research wil be reviewed.
Definitions
Biomarker(1)
“A defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention, including therapeutic interventions. Molecular, histologic, radiographic, or physiologic characteristics are types of biomarkers. A biomarker is not an assessment of how an individual feels, functions, or survives. “
Examples
• susceptibility/risk biomarker
• diagnostic biomarker
• monitoring biomarker
• prognostic biomarker
• predictive biomarker
• pharmacodynamic/response biomarker
• safety biomarker
Radiomics(2)
“conversion of digital medical images into mineable high-dimensional data… motivated by the concept that biomedical images contain information that reflects underlying pathophysiology and that these relationships can be revealed via quantitative image analyses”
Examples of Imaging Features
There are multiple possible grouping of imaging features that can be used as IB’s. For example, 1st order features are based on each pixel value independent of surrounding pixels and provide information related to histogram of the image. Second order features are based on the spatial relationships among the neighboring pixels (gray-level configuration); they provide information about the texture of the image. Groupings may also be done by contour or shape features, size, texture etc.Segmentation
Segmentation to derive IB’s can affect the precision of IB’s. There is an increasing awareness of the need to provide semi-automated or automated segmentation algorithms to radiologists to help with the validation of IB’s as many potential IB’s struggle with poor reproducibility. Machine learning and Deep Learning analytic approaches have strong momentum in this area and a growing role in the image analysis pipeline of IB’s in general achieving human or better than human performance in non-medical imaging domains and exhibiting promising initial results in simple medical imaging segmentation and interpretation tasks.Statistical Considerations/Shared Imaging ArchivesA common problem encountered in radiomics is loss of statistical power related to multiple measurements in a single patient across multiple tumor sites. Thus, strategies for variable reduction must be applied to avoid the need for large clinical trials with thousand of patients. Drawing from previously published work where IB’s have already been established is one approach. Combining IB’s into a single IB signature and then retesting it in another cohort is another. Searching for correlations and selecting representative features from each correlation cluster is another. A combination of these approaches is often used. Imaging data sets in the public domain can be used either for validation or hypothesis generating analysis and derivation of IB’s. The Cancer Imaging Archive (TCIA) is an example of an example of one data set repository useful for validation and early development of IB’s (http://www.cancerimagingarchive.net/)The Discovery and Validation of an Imaging Biomarker (IB)
The discovery and validation of an imaging biomarker should follow a well defined roadmap for translation. This has been outlined well in a recent consensus paper by O’Conner et al (3). Some key questions must be answered at an early stage: “Does the IB fulfill an unmet clinical need?”; “Does data exist to evaluate the IB and if not can it be obtained?”. At an early stage, technical validation including assessment of precision through repeatability and reproducibility must be determined. Furthermore, biologic and clinical validation must also be performed. Cost effectiveness must also be considered. These steps are usually done in a single center in a smaller cohorts.Once completed the same paradigm needs to be repeated to some extent in a multi-institutional setting and finally the IB validation will proceed on to multi center prospective trials. These qualification steps can be applied in multiple contexts including screening IB, diagnostic IB and predictive IB.Statistical Considerations/Shared Imaging Archives
A common problem encountered in radiomics is loss of statistical power related to multiple measurements in a single patient across multiple tumor sites. Thus, strategies for variable reduction must be applied to avoid the need for large clinical trials with thousand of patients. Drawing from previously published work where IB’s have already been established is one approach. Combining IB’s into a single IB signature and then retesting it in another cohort is another. Searching for correlations and selecting representative features from each correlation cluster is another. A combination of these approaches is often used. Imaging data sets in the public domain can be used either for validation or hypothesis generating analysis and derivation of IB’s. The Cancer Imaging Archive (TCIA) is an example of an example of one data set repository useful for validation and early development of IB’s (http://www.cancerimagingarchive.net/)Acknowledgements
No acknowledgement found.References
1. Group F-NBW. BEST (Biomarkers, EndpointS, and other Tools) Resource [Internet]. BEST (Biomarkers, EndpointS, and other Tools) Resource. 2016. Available from: http://www.ncbi.nlm.nih.gov/pubmed/27010052
2. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology [Internet]. 2015;278(2):151169. Available from: http://pubs.rsna.org/doi/10.1148/radiol.2015151169
3. O’Connor JPB, Aboagye EO, Adams JE, Aerts HJWL, Barrington, Sally F, Beer AJ, et al. Imaging Biomarker Roadmap for Cancer Studies. Nat Rev Clin Oncol [Internet]. 2016;in press. Available from: http://dx.doi.org/10.1038/nrclinonc.2016.162
Figures
