The Future of Radiomics

The field of radiomic biomarkers seems to be maturing with higher levels of evidence being generated through validation in randomized trials. However, these biomarkers, currently, are inductive - based on a retrospective training set, we hope the signature/model extrapolates to a new patient. It only takes one black swan (aka an unknown unknown) for the model to crash and burn. This is particularly a huge problem with deep learning approaches and I am therefore more bullish on hand-crafted radiomics. Within radiomics, I believe shape features are much more interpretable than texture features.

I see the way forward as radiomic biomarkers being incorporated in the standard biomarker discovery process in earlier phases of trials. This will largely be a collaborative effort between pharmas and small/medium companies with expertise in the AI/radiomic side. However, the key point is that while prospective validation may be convincing, it might not provide sufficient immunity to black swans. Sure, one will likely make a decent amount of money in the short term (MBA types are probably excited here) - since by nature black swans are extremely rare events. In the long term, such an occurrence may set the field back significantly.

Pre-clinical studies of radiomics through micro-CT scans or the like will allow us to more accurately "explain" the disease processes captured by radiomic features driving the model. And good explanations in turn enable better predictions. In summary, here's my long term outlook. We've already completed or are currently in the phases highlighted in italics. Each additional step will be exponentially more costly, time-consuming and difficult. But hey, when did we stop dreaming big? 

Discovery using real-world institutional data > Multi-institutional validation > Validation in completed randomized trial(s) > Prospective observational validation > (optional) Prospective interventional validation > Integration into standard biomarker discovery process for new therapies > Mechanistic explanation of factors driving radiomic predictions

Interestingly, this is in contrast to the typical drug development process which follows a bottom up approach:

Mechanistic explanation of disease pathology > Pre-clinical studies of potential modulators > Prospective human trials > Post-market studies and follow-up reporting of completed trials > Real-world data generation in patients at institutions