Radiomics and Pathomics as Emerging Fields
Currently, radiologists and pathologists can only report what they see in individual medical images, but doing so “only captures a fraction of the information in the image,” Daniel Rubin, MD, professor of biomedical data science, radiology and medicine at Stanford University in California, told Cancer Therapy Advisor. “There’s a lot of information the human eye doesn’t recognize and can’t appreciate, but a machine can. A pathologist can’t look at every single pixel, and even if you can, there’s a lot that a computer can see that you may not be able to see.”
But there’s a caveat to seeing so much — the risk of seeing too much.
“What the computer sees may not be real or relevant or reproducible across scanners or patients, so just because you see more doesn’t mean it’s more that you want to see,” Dr Rubin said. “The issue, from the research point of view, is to figure out what is a real signal that is actionable versus what might you see that isn’t anything to which you need to pay attention.”
To be clinically useful, radiomics and pathomics require data from huge numbers of images — at least tens or hundreds of thousands. From these large collections of images, researchers can use radiomics methods to pull “quantitative features and look for similarities in those features across a population of patients who otherwise appeared similar by human inspection,” Dr Rubin explained. Those features comprise a digital signature, or an electronic phenotype, for characterizing the disease. Machine learning with radiomics provides methods for determining and validating electronic phenotypes using historical imaging data. Machine learning with pathomics works the same way, but uses images of tissues.
In their paper, Dr Banna and colleagues succinctly summarized the steps in this process:
- Acquisition of images
- Identification of the volumes (habitats)
- Segmentation of the volumes (contouring)
- Image features extraction (quantitative)
- Analysis (correlation).
Limitations of Existing Single Biomarkers
The hope and expectation is that these electronic phenotypes, representing different subtypes of cancers and people, can overcome the limitations of individual biomarkers seen in recent clinical trials.
Take, for example, TMB in the KEYNOTE trials. At the 2019 World Conference on Lung Cancer, researchers reported that tissue TMB (tTMB) did not predict the efficacy of combining chemotherapy with pembrolizumab in the KEYNOTE-189 trial.2 At the same conference, findings from KEYNOTE-021 similarly disappointed, leading Charu Aggarwal, MD, MPH, a lung and head/neck oncologist and assistant professor at Penn Medicine in Philadelphia, Pennsylvania, to ask on social media, “Is TMB dead? Maybe plasmaTMB will have a different tale to tell.”3
But the handicap plasmaTMB shares with tTMB is its singularity.
“Any kind of biomarker, whether it’s imaging-based or molecular, has a limited amount of applicability,” said Fred Prior, PhD, chair of the department of biomedical informatics at the University of Arkansas for Medical Sciences and a researcher at the Winthrop P. Rockefeller Cancer Institute in Little Rock. “Any marker that’s going to be reasonably viable is, first, only going to work on subpopulations, and second, not going to be single parameters but combinations of parameters that together give us more specificity.”