NEW YORK – Picture Health is hoping that the broad therapeutic applicability and interpretability of a new AI-derived biomarker test based on tumor blood vessel twistedness will help it advance the tool rapidly to commercialization.
The biomarker, dubbed quantitative vessel tortuosity (QVT), has been shown in recently published studies to predict how patients with different types of cancer respond to various therapies, including immune checkpoint inhibitors. The biomarker evolved out of the observation that as tumors grow, the architecture of the vessels supplying blood becomes twisted and disorganized, fostering a treatment-resistant tumor microenvironment.
Picture Health CSO Anant Madabhushi said the idea of using blood vessel tortuosity as a biomarker first occurred to him over 20 years ago, while he was in graduate school. "I saw a poster about brain tumors and how blood vessels were different in benign brain tumors versus malignant brain tumors, and I found that really interesting," Madabhushi said.
Later, when he was a faculty member at Case Western Reserve University, he returned to the idea during a study using AI and machine vision technology to try to better distinguish whether a lung nodule was malignant or benign. On CT scans, radiologists have difficulty distinguishing lung nodules, called granulomas, from malignant adenocarcinomas.
"The fact that this was very difficult to distinguish based on just morphologic or routine image features of the CT scan was one of the reasons why we started to look at the twistedness of the vasculature in the context of lung nodules," Madabhushi said. With the help of AI, the researchers found that benign nodules had much smoother vasculature than adenocarcinomas, which had very twisted, convoluted vasculature.
In a paper published last October in Clinical Cancer Research, Madabhushi and colleagues showed using routine pre-treatment radiology scans from 558 cancer patients receiving first-line therapy that QVT measurements could predict how patients responded to treatment. They further found that QVT-based risk scores were prognostic and could determine recurrence in patients with different cancers and on a range of treatments.
In another study, published the following month in Science Advances, Madabhushi and colleagues concluded that tumor vessel tortuosity in metastatic non-small cell lung cancer patients was associated with response to immune checkpoint inhibitors and prognostic of overall survival. QVT was also predictive of response and prognostic of survival in patients receiving first-line chemotherapy with anti-PD-L1 immunotherapy. And in patients with early-stage tumors who had never received immune checkpoint inhibitors, higher QVT was strongly associated with PD-L1 expression.
Another interesting finding when comparing the tumor vasculature of patients pre- and post-treatment, Madabhushi said, was that the tumor vessels got more twisted in patients who did not respond to immunotherapy. Conversely, the same comparison in patients who responded to immunotherapy showed tumor blood vessels were less twisted than they were at baseline.
This suggested to Madabhushi that QVT could be used clinically to predict patients' responses to immunotherapy. That vision may become reality now that Picture Health has committed to developing the biomarker into a commercial test.
Cleveland-based Picture Health, a startup that Case Western Reserve University launched last year, licensed the patent portfolio for this technology from the university. To facilitate commercialization, the company is assembling larger datasets and combining the blood vessel tortuosity measures with other features, according to CEO Trishan Arul.
"We have our feature library that correspond to treatment outcomes, and what we want to do is make a more complex, more generalizable algorithm, which will look not just at QVT, but also these other features, enabling oncologists to then determine what the best course of treatment would be for a patient."
Picture Health is in discussions with a number of pharmaceutical companies about commercialization plans and is working with some drugmakers to integrate QVT and these other imaging features within ongoing clinical trials. Arul declined to disclose these pharma partnerships, but said that the company's goal is to ultimately develop QVT into a companion diagnostic.
Picture Health sees potential for QVT to be paired with genomic and protein tests used to direct cancer therapy today. As an example, Madabhushi noted that in the Science Advances study, researchers homed in on a subset of patients with low PD-L1 expression who received combination chemotherapy and an immune checkpoint inhibitor, and using the QVT biomarker, researchers were able to parse out which patients would likely do well with just an immune checkpoint inhibitor. "It represents a limitation of the PD-L1 biomarker, and that currently is the gold standard for deciding what kind of therapy you're getting as far as immunotherapy is concerned," said Madabhushi.
Arul said that Picture Health chose the QVT biomarker out of its portfolio of technologies to move forward right away because of its applicability across multiple treatments and cancer types. Another appealing quality of QVT is its intuitive interpretability, which helps build trust among oncologists in AI-generated recommendations.
"When you look at the images, it becomes very obvious as to what a responder is and what a non-responder is, because you can appreciate immediately the twistedness of the vasculature associated with the tumor," said Madabhushi.
The fact that oncologists can easily understand the scientific rationale behind QVT's therapy-predictive capabilities also gives Picture Health confidence in deciding to commercialize this test. "I call it the head nod validation," Madabhushi said. "Explain the idea of vessel tortuosity, and you see heads nod. It means people understand what this biomarker is about." That's in contrast with the typical "black box" approach to AI, where predictions are produced using methods that are unknown to the operator.
That issue has been part of a larger debate as AI technologies expand beyond early applications like image interpretation, which can be verified by a human expert. "As we get into the precision oncology space where we're now saying here's how you should treat a patient, you can't do this with your naked eye," said Nathaniel Braman, Picture Health's AI director, "then, [interpretability] becomes critical because we're asking a lot more of oncologists. We're asking them to make a decision based on something that they can't spot check themselves."
QVT's interpretability will undoubtedly come up in regulatory discussions when the time comes for Picture Health to seek the US Food and Drug Administration's approval for a test using this biomarker. With AI-based or imaging-based algorithm tests, the FDA will typically want to compare the new technology to what a human does, Arul said, but with QVT, that's not possible.
Since the agency has never approved this type of AI algorithm before, the company will have to break new ground in its interactions with the FDA. "One of the big challenges we're dealing with now is working through the process with the FDA and figuring out what would the validation study look like," Arul said. "Our regulatory consultants are tackling that challenge right now."
Braman said that moving forward, some challenges Picture Health will face commercializing the QVT biomarker include eliminating potential sources of bias in the results and making sure the test performs consistently across different CT and MRI scanners.
For the patient, implementation of the test would be fairly seamless, according to Arul. "We're not expecting anyone to order new scans," he said. We're expecting to be able to run our AI algorithms on the existing scans that they have."