Editorial: Radiomics: The New World or Another Road to El Dorado?

The emerging field of radiomics is generating remarkable interest from oncologists and cancer researchers alike. In the current issue of the Journal, the commentary by Verma et al. (1) describes the future applications of radiomics, highlighting its seemingly limitless potential while providing perspective on the methodological challenges facing the emerging field. To this end, the paper successfully introduces radiomics as a new “vein” of exploration for personalized cancer medicine. While this is an exciting and promising perspective, it fails to highlight the underlying methodological and conceptual shift in cancer research that radiomics represents and in doing so unintentionally highlights one of the greatest problems with cancer research today—our fascination with finding the road to El Dorado. While the importance of computer-aided diagnosis (CAD) and decision support systems for improved medical care was recognized almost 50 years ago (2,3), technological advancements over the past 20 years have dramatically increased the number and variety of measurements being made in the processes of delivering care and in performing clinical research. To be clear, this is not simply a reference to genomics—the increases are now coming on all fronts, including volumetric energy-selective computerized tomography, combined magnetic resonance–positron emission tomography scanners, anesthesia machines that stream data to the cloud, and personal health monitoring through devices like the Fitbit. The volume, variety, velocity, and veracity of measurement data collected per patient, including traditional prognostic and predictive factors, has led to a data aggregation and analysis task that frankly exceeds human cognitive capacity (4) and puts patients and clinicians in the situation of suboptimal use of the costly and risk-associated measurements that have been collected in the process of delivering care. Furthermore, the paralysis induced by our limited cognition and our isolation of data prevents us collectively from learning as much as possible from every patient. It is unclear if the addition of another “signal” in the form of radiomics will make this situation better or worse. We would argue that the major contribution of radiomics is not in the identification of a set of image-based biomarkers, but rather in its illumination of a critical “tipping point” in technological capability that will both challenge and enable the clinical and research oncology community. After all, the process of automatic conversion of medical imaging data to quantifiable features that can be mined and used for objective and automated decision support (5) was introduced in the 1970s by Hall et al. (2) and Harlow and Eisenbeis (3). It was more than 40 years later that the real value of image feature quantification for decision support was reported and coined as “radiomics” (6). While interest in the field has since skyrocketed, it is important to highlight what is really important here—what is the critical novelty in every impactful radiomics publication? We would argue that it is not in the identification of yet-another-‘omic (YAO) signal that clinical scientists can cast their patient outcome measures against. Rather, the true impact is in their methods—they employ pipelines of automation to analyze otherwise discarded radiological imaging data. Radiomics is exciting because it gives us a glimpse of the future way we will do science—like it or not. The development of highly automated analysis pipelines is driving pharma, industry, and academic health centers to make massive investments in “data lakes” capable of ingesting and storing features regardless of their origin—volumetric imaging, deep sequencing, detailed treatment records, and patient outcomes, to name a few. Completely automated feature extraction techniques are mature and operating at scale today in platforms such as Google’s image search. Such feature extraction and analysis methods are readily distributed across extremely diverse source data, allowing investigators to take broader perspectives and to explore de novo correlations. This is reinforced in the work of Aerts et al. (7), wherein genomic, computerized tomography image texture, and tumor shape were all included in the development of predictive signatures. The ongoing development of artificial intelligence (AI) approaches that employ deep learning methods will accelerate these efforts and set about “automatically starting the automation” to search for patterns and correlations not yet explicitly considered by humans. Heretofore, the value of such “deep automation” could be considered nonscientific and would often be deemed as lacking with respect to the underlying hypotheses being tested. Radiomics’ use of otherwise discarded imaging data makes a profound statement regarding our hubris in setting bounds on the validity and acceptance of alternative methods of advancing science. The team that has advanced the field of radiomics has also been building a global network of distributed machine learning “bots” that don’t even require the data to be in one place (8). Such technological advances, including the massive investments being made in AI technology, portray a world in which humans and machines work “collaboratively” across masses of distributed data to accelerate progress in our fight against cancer. In this new world, we will need to be vigilant in pursuing a deep understanding of the underlying mechanisms of action—these will contribute most directly to the development of biologically targeted cancer interventions, optimization of existing therapeutics, and personalization of cancer medicine. In this new world, the traditional methods of scientific pursuit, data sharing, and discovery will be challenged and the rate of progress against cancer will be accelerated—that is exciting. The authors have no conflicts of interest to disclose.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.