Skip to main content
Premium Trial:

Request an Annual Quote

Picture Health Shows Potential of AI Imaging to Guide Cancer Treatment

Premium

NEW YORK – Researchers from Case Western Reserve University have launched a startup company based on computer vision technology that successfully predicted which cancer patients would benefit from treatment with an immune checkpoint inhibitor in a lung cancer clinical trial.

The company, Picture Health, licensed technologies developed within the Center for Computational Imaging and Personalized Diagnostics at Case Western for risk stratification and treatment response prediction for cancer patients. Picture Health Cofounder Anant Madabhushi spearheaded the research involved in the agreement, which uses artificial intelligence to identify patterns in routinely acquired data, such as CT scans and pathology images, to create an imaging biomarker associated with response to cancer treatment, including modalities such as immunotherapy. Picture Health will now pick up the baton from Case Western to translate and deploy that technology to a clinical setting.

One of the differences between Picture Health and other startups in the imaging AI space is that it has some evidence behind its algorithms, according to its founders. "We're not just starting from a bunch of algorithms," said Madabhushi. "A large part of the retrospective validation has been done. We're starting, not from the ground floor but probably the fourth or fifth floor."

That's reflected in a recent publication in NPJ Precision Oncology from the Case Western group describing the development and validation of a predictive pattern of tumor-infiltrating lymphocytes in non-small cell lung cancer using a combination of images from various sources.

Pairing pathology slide images with patient outcome information from the Cancer Genome Atlas, Case Western researchers used machine learning to model morphologic and molecular differences in immune patterns between the two most common forms of NSCLC — lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). In their training set, they found that the density signature of TILs (DenTIL) was prognostic for overall survival in LUAD, and that the spatial distribution signature of TILs (SpaTIL) was prognostic of overall survival in LUSC.

When tested on an independent data set also drawn from slide images from the Cancer Genome Atlas, and on an external validation set of images from 123 patients with locally advanced LUAD from the University of Bern in Switzerland, the DenTIL measure was a statistically significant prognostic indicator. Because the validation set included patients treated with a number of different neoadjuvant chemotherapy regimens, the researchers concluded that the DenTIL signature is able to capture hallmarks of disease aggressiveness across different tissue types and under various treatments.

They then conducted another validation experiment using data from Bristol Myers Squibb's Checkmate-057 Phase III clinical trial of Opdivo (nivolumab) versus docetaxel in patients with advanced non-squamous NSCLC who had previously failed platinum-based chemotherapy. Results from this second external validation showed that the DenTIL signature was predictive of response to treatment with nivolumab but not docetaxel. In this data set, DenTIL was not prognostic of overall survival for either arm of the study.

Building on that validation of its algorithms, Picture Health will move forward in lung cancer as its initial area of development with the ultimate goal of expanding across all tumor types. "We're a small startup," said Picture Health CEO Trishan Arul. "We're obviously going to start in one place as opposed to trying to boil the ocean. But we expect to quickly move into other types of solid tumors."

In 2020, AstraZeneca and BMS agreed to provide the Case Western computational imaging group with imaging data from completed clinical trials in which immunotherapy drugs were tested on lung cancer patients. While those agreements won't directly carry over to Picture Health, the company will look for further opportunities to work with pharma companies to bring its technology to patient care. "We're already in discussions with a half-dozen potential partners," said Arul. "For example, the development of a companion diagnostic that will help to segregate patient populations into those who are likely to respond to treatment versus those who won't respond to treatment."

Picture Health joins a growing space in AI-based image analysis for cancer that includes established firms like Paige and PathAI, as well as newcomers like Lunit and Owkin. In work closely aligned with Picture Health's, Seoul, South Korea-based Lunit recently published proof-of-concept research showing its AI platform can predict response to checkpoint inhibitor therapy based on an analysis of TILs in tumor tissue. Its classifier was particularly useful in patients with lower PD-L1 expression, between 1 and 49 percent, identifying patients in that group who could derive benefit from checkpoint inhibitors.

Meanwhile, Owkin bagged a $180 million equity investment from Sanofi in November 2021 to support discovery and development programs in NSCLC, triple-negative breast cancer, mesothelioma, and multiple myeloma. The collaboration covers identification of new biomarkers, building prognostic models, and predicting response to treatment from patient data. And the Paris-based company inked another deal last week worth up to $180 million with Bristol Myers Squibb to apply its machine-learning technology to improve the design and execution of clinical trials in cardiovascular and other diseases.

There are several factors differentiating Picture Health from other companies in the space, however. "We believe we're adding a value proposition to clinical problems that doesn't exactly seem to be addressed by these other competitors," said Madabhushi. "And our technology is straddling multiple different modes of data," meaning data from both pathology and radiology.

One of those differentiators is that Picture Health's approach is interpretable. "We are the opposite of the black box," Madabhushi explained. "One of the challenges with a lot of deep learning approaches is that they are very opaque, because it's unclear what the machine is actually learning, what the machine-learning algorithms are, or what the nature of the representations being learned from the data is that is then associated with the outcome or diagnosis."

Madabhushi and his colleagues took a deliberately contrarian approach in developing technology that creates a feature library that can be studied in relation to the underlying biology of the cancer.

"If you look at the whole plurality of clinical problems in the context of cancer care, we're able to go in and look in our feature library and identify that subset of features or patterns that can allow us to address specific problems," Madabhushi said. For example, "trying to predict the response to a therapy for a given patient with a particular kind of lung cancer."

In addition to working with pharma companies to support drug development in clinical trials, the Picture Health founders plan to develop their technology for clinicians to use for patient care directly. What that would look like in practice would involve integrating Picture Health's software into existing picture archival and communication (PAC) systems, which are typically already set up to work with an AI provider. Images would be uploaded from the PAC into Picture Health's cloud-based system, which would analyze them on demand and send them back. For PAC systems that aren't set up to integrate with an AI provider, a manual upload alternative would be available.

The range of images would include those generated by either radiology, such as CT or MRI images, or pathology, like hematoxylin and eosin stained or immunohistochemistry slides. "We're able to use those images that already exist and are already routinely acquired," said Arul. "We're not asking anyone to do anything new. Our goal is to integrate with various PAC systems so that we're able to pull through the images, do the AI analysis, and provide back a report." And the software would be enabled to provide reports to the radiologist, pathologist, and oncologist, not just a single provider on the care team.

In terms of the impact of that analysis, Madabhushi said that it would be focused on predicting and monitoring patient treatment response.

"Physicians really don't have very good tools at their disposal to be able to predict which patients are going to respond to certain therapies or certain combination therapies upfront," said Madabhushi. "And then, critically, even after they have started getting treated with a particular drug or drug combination, it's very difficult to tell with the current metrics that are being employed whether the treatment is working."

Arul and Madabhushi also envision Picture Health's system as a tool to "add granularity" to treatment decisions for oncologists, and that would include making predictions about therapies linked to genomic biomarkers. In one example of the kind of questions Picture Health plans to pursue, Case Western researcher Xiangxue Wang and colleagues used the same AI software that is now licensed to Picture Health to investigate outcomes in patients being treated with checkpoint inhibitors. That study showed that the AI tools could actually stratify clinical outcomes in patients with low PD-L1 expression, suggesting that this category of patients could avoid more aggressive chemotherapy and might benefit from immunotherapy alone.

Fine tuning treatment recommendations for the use of immune checkpoint inhibitors alone or in combination with other therapies is compelling enough that the FDA presented two studies at the American Society of Oncology's annual meeting in Chicago earlier this month. In one, the FDA retrospectively explored outcomes in NSCLC patients treated with anti-PD(L)1 checkpoint immunotherapy alone or in combination with chemotherapy in a first-line setting, finding patients receiving combination therapy had an edge over those that received immunotherapy alone. A second, similar analysis focused on KRAS mutations, showing that advanced NSCLC patients with both KRAS-mutated and KRAS wild-type tumors had increased overall survival times and objective response rates when they received chemo and immunotherapy as a first-line treatment.