We recently interviewed Nathaniel Braman, Co-Founder and Head of AI at Picture Health, a precision medicine company that applies novel AI algorithms to radiology and pathology medical imaging. See our takeaways here, and the full transcript below.

‍

‍Key Takeaways:

• Radiomics holds untapped potential for pharma clinical trial biomarker strategy to improve tumor profiling, patient characterization / stratification, and early response detection, leveraging routinely collected data that does not require additional sample collection, creating opportunity to streamline clinical development timelines and realize cost savings
• Radiomics fills gaps in current pharma biomarker discovery and development by better addressing phenotypic assessments, playing a complementary role with other tech modalities
• There are two main camps of radiomics approaches, each with unique strengths and weaknesses. Deep radiomics / deep learning approaches are powerful prediction tools; however, they have huge data requirements and little explainability. On the flip side, traditional feature-based radiomics are much more data efficient and explainable, but can often fall short of providing direct mechanistic insight. Nate describes Picture Health’s ‘biologically inspired radiomics’ approach to marry the two
• While many pharmas are recognizing the potential of radiomics, adoption is moderated by data dependence, evolving ​trust, interpretability​ of these tools, and regulatory ambiguity

‍

‍

Thank you for joining us today! To start, could you share a little bit about your background and Picture Health?

In 2021, I co-founded Picture Health with my former thesis advisor, Anant Madabhushi. Before that, I’ve worked in the radiology AI and precision oncology spaces at IBM Research and Tempus. Picture Health’s mission is to transform cancer care and drug development using standard-of-care clinical imaging data, such as radiology and digital pathology.  

As we began building Picture Health, we spent a lot of time looking at how radiomics was being used in practice. We saw potential of AI and medical imaging to deliver a new generation of biomarkers for guiding precision oncology. But we also saw a disconnect between current radiomics approaches and the kinds of biomarkers that clinicians and drug developers trust and use today. There seemed to be two main camps in the field. On one side, there’s the “Deep Radiomics” approach, where large neural networks are trained to directly predict outcomes or guide treatment decisions. These models can be powerful, but they require training datasets of hundreds to thousands of patients - a challenge in the clinical development setting - and their predictions are opaque and difficult to interpret. On the other side, feature-based radiomics uses engineered imaging measurements - known as “features” - as inputs to more lightweight machine learning models. These models are more data-efficient and transparent, since their outputs can be traced back to specific features. But the standard feature sets that are commonly used aren’t much better in terms of biological interpretability. Knowing how something is calculated doesn’t necessarily tell you what it means for tumor biology or patient care.

So, at Picture Health, we merge these approaches by building AI imaging biomarkers that can not only improve precision medicine and drug development but also help us better understand cancer biology. We focus quite strongly on what we call ‘biologically inspired radiomics.’ We use deep learning tools to map tumor burden throughout the body, as well as the tumor habitat. From there, we compute our library of radiomic feature classes, each of which are designed to capture a critical aspect of tumor biology.​ In this fashion, radiomic tools could be used to solve critical challenges in clinical trials, while also providing interpretable insights related to an agent.​ ​     ​

‍

Pharma often sits in an interesting position, where they have the power to push new technologies into the clinic. Thinking about biologically inspired AI-based radiomic biomarkers, can you give us some examples of pharma use cases?

We know biomarkers are critical to high-stakes oncology clinical development. A non-biomarker clinical trial is twice as likely to fail as one that has a biomarker. Obviously, we want every trial to have biomarkers. However, traditional blood and tissue-based biomarkers are often subjective, costly, susceptible to biopsy sampling error, and may often be tissue-destructive.

In contrast, radiomics makes use of scans that are already routinely collected and addresses several other limitations of tissue-based biomarker strategies: it can provide non-invasive, in vivo characterization of multifocal disease comprehensively throughout the entire body. Because of these strengths, radiomic biomarkers can have significant impact across the entire lifecycle of an asset, from Phase 1 through 4. We think of this in terms of signal and see opportunities to accelerate clinical development by finding, generating, enhancing, and utilizing signals in drug development.

As a tangible example, we’ve built a biomarker we call QVT Phenotype, centered on one of our biologically inspired feature classes: quantitative vessel tortuosity, or “QVT.” For QVT, we automatically map out both tumors and their surrounding blood vessels using AI, then measure the shape and organization of the tumor-associated vessel network. Using this approach, we’ve discovered that there are intrinsic phenotypes of vascular complexity that stratify patients by treatment response, even before therapy begins. Patients with QVT Phenotype corresponding to complex vasculature - indicating unchecked tumor angiogenesis - tend to have poor therapeutic outcomes.  

We’re showing some exciting results at ASCO this year of a validation study demonstrating that this tool can predict outcomes to multiple ICI-based regimens, but we’ve also shown that QVT Phenotype is prognostic across a range of tumor types and therapeutic classes - not just immunotherapy. You can use radiomics to better help characterize patients, even before treatment, based on their likelihood of response or resistance​ in a way that is aligned with known tumor biology. ​  

​​On top of this, a further significant advantage of radiomics relative to conventional biomarker approaches is that it’s longitudinal. While a tissue-based biomarker is a snapshot in time, a radiomic biomarker is dynamic because it can track the evolution of treatment response. ​For example, we’ve ​been able to effortlessly deploy QVT Phenotype as a longitudinal tool to monitor how tumor angiogenesis evolves. This is where we start to generate signal, particularly in the early trial setting. For instance, when a patient’s QVT Phenotype shifts from high to low complexity after initial treatment, that change ​indicates treatment-related vascular normalization and ​is an early indicator of therapeutic benefit.​ ​​​​​     ​Because we can generate early mechanistic signals in Phase 1/2, we can provide evidence to support dose selection and early efficacy evaluation far earlier than standard endpoints allow.

Further down the pipeline, ​     ​​you can apply these tools ​to enhance signal within trials. For example, we’ve found these ​QVT ​phenotypes to be prognostic independent of conventional biomarkers such as PD-L1 and clinical endpoints like RECIST. And when we integrate them, we find that we can better stratify patients by outcome and uncover treatment-responsive subgroups that might be missed otherwise.

You can also utilize this tool in the late-stage trial setting by enabling adaptive designs or supporting biomarker-driven regulatory strategies. In some cases, we’ve seen this tool help refine inclusion/exclusion criteria or support exploratory endpoints in registrational trials.

​​​

‍Thanks for the QVT case study! ​​This is an interesting example of emerging classes of radiomic biomarkers.

A lot of the early precision medicine work leveraged genomic or IHC-based tests, for example. How do you see radiomics fitting in this broader landscape of tools, and how does biopharma navigate this range of tools?

One of the reasons radiomics is so attractive to drug developers is that it addresses several limitations of tissue-based, molecular biomarkers. The traditional biomarker discovery pipeline is notoriously slow and costly. Molecular biomarkers often face significant hurdles such as limited sample availability, invasive collection procedures, development effort and cost, and reproducibility.

But radiomics makes use of standard clinical imaging, which is already a cornerstone of clinical trials. It’s non-invasive and non-tissue destructive. It can measure spatial and temporal tumor heterogeneity across the entire body, capturing disease dynamics that are often missed by tissue-based sampling. Radiomic biomarkers provide entirely new dimensions of tumor characterization, with higher accessibility. Because of these strengths, it’s possible for imaging biomarkers, built using radiomic features, to be rapidly prototyped, validated, and deployed throughout a drug’s development cycle of a drug in a way that is far more challenging for molecular biomarkers. ​But it may require data efficient AI approaches to make use of the limited datasets that are available early in a drug’s lifecycle.  ​​     ​

What excites me most, though, is how well these two technologies can complement each other. Cancer is a multifaceted, multiscale disease. Integrating radiomics with molecular biomarkers allows us to capture both phenotypic and genotypic dimensions of cancer, creating a more comprehensive and precise biomarker strategy. By bringing these tools together, ​you​​ ​can characterize treatment mechanism-of-action across scales, which is going to result in stronger therapies and better patient outcomes.  

‍

What is the opportunity for imaging biomarkers beyond radiology?

Imaging biomarkers are important because they allow us to understand the phenotype of the tumor and microenvironment. Critical biology related to treatment response and resistance can not only be measured but spatially mapped - providing greater insight into tumor heterogeneity. When you look at a biomarker like PD-L1, which has notoriously inconsistent predictive performance across tumor types and indications, it’s clear that traditional tissue-based biomarkers are not sufficient to capture the vast complexity of treatment response and resistance. We need new angles and fresh tools.

Digital pathology is another exciting emerging biomarker frontier, and it’s also part of our arsenal at Picture Health. It gives us the ability to really zoom in and quantify the tumor and tumor microenvironment at the microscopic scale. You can, for instance, detect and measure subtle interactions between tumor cells and immune cell subpopulations to reveal the complex patterns of immune response or the orientation of collagen fibers to understand the complexity of the extracellular matrix.​ Interestingly, radiology and digital pathology tend to bring different, complementary information to precision medicine problems.​  

‍

Where do you think we are in terms of the adoption curve? What are the use cases that you see pharma exploring today, versus which are less proven or are further out?

​​It feels like we're rounding the top of the first hill of the roller coaster. At this point, most, if not all, global biopharmas are recognizing the underutilized value of their imaging data and they're curious where AI can fit in. Many are excited about how AI-augmented workflows can improve efficiency, for instance in response evaluation. Recently, we have partnered with Friends of Cancer Research and a few major pharma players on a working group to tackle this challenge, and they’re excited by the potential of AI to streamline their centralized review process and reduce variability of human readers. ​​​

Increasingly, pharma companies are hiring their own teams of AI experts, who can speak this new language and interface with third party developers. Among those, there’s a smaller, but growing, group of biopharma, who are pushing the boundary of how these tools can shape the future of clinical trials, for instance as​     ​ AI imaging biomarkers. They are connecting the dots from radiomics to more efficient patient screening, earlier decision-making, and stronger survival improvements, and they are putting these ideas to the test. ​     ​There’s an exciting feedback loop that happens here: integrating radiomics into clinical trials for these purposes can improve the success of clinical trials, which in turn pushes these tools towards clinical practice. I could imagine a future where your doctor could be consulting a radiomics report the same way that they do a sequencing or a molecular biomarker report. But it all starts with pharma.  

‍

What are the current challenges that tools developers such as yourself encounter? And on the flip side, what are challenges that pharma are navigating?

Starting with developers, AI always has a dependence on data, and that challenge is significantly amplified in this space because we're dealing with very specific subsets of patients. So, data scarcity is something that you must overcome if you want to build a tool for a novel treatment or a particular treatment niche. A core problem that developers need to solve is data efficiency. If you're going to keep up with emergent therapies, you must be able to work with small data sets to build tools very quickly.  

Another huge challenge for developers is trust and adoption. This is a very new idea, and there's a lot of education required. We need to be teaching pharma collaborators, the FDA, and clinicians how this technology works, how it can be used, and that it's not some scary black box AI. It can be understood and biologically intuitive, something akin to a molecular assay, except that it comes from imaging. This is why we focus so strongly on the interpretability piece.  

Pharma faces this challenge too. They want the best technology and the strongest biomarkers. But there’s always new tools. How do you choose the right ones to invest in?  The regulatory ambiguity contributes to this. This is a very new idea for the FDA, using images in this fashion. It's a totally new frontier. If you're serious about getting these tools into your trial, you need to have a plan to bring these ideas to the FDA, get them educated, and all of that.  

But there’s been great progress here. As uncertainties fade, these challenges become a strength. The promise of radiomics is so large precisely because it is so distinct from typical biomarkers. It can give you a completely new perspective.    

‍

Looking ahead, what developments are you most excited about?

There’s much to look forward to. I’m most excited about seeing these tools finally take a more central role in clinical trials. I’ve been working to advance this technology for a decade now. For much of that time, radiomics felt like it was stuck in an exploratory space - interesting science, but not yet actionable. Now we’re seeing a real shift. We’re working with pharma partners who are not only open to using these tools but actively looking to integrate them in ongoing trials. That’s huge.

Radiomics is already being used to inform inclusion criteria, to track early response, and to identify non-responders before treatment even begins. Radiomics is an emerging valuable tool in the evolving landscape of cancer treatment, with several areas under active investigation. And all of this raises a bigger question: once a drug is approved, how do we keep using these tools? Can they support companion diagnostics? Could we deploy them in real time to validate and apply predictive biomarkers across the drug’s entire lifecycle? The potential to follow a biomarker from discovery all the way through to market gets me really excited.  

And of course, all of this ultimately leads back to patients. Can radiomics bring more effective treatments to market? Can radiomics find the treatment that is going to make the difference when a tumor just won’t respond to anything else? Can it give patients more years with their family? I’m looking forward to finding out.  

‍

Thanks so much, Nate, for your insightful perspectives.