Skip to main content

Radiomics Could Complement Molecular, Clinical Data in Cancer Care, Study Finds

Premium

NEW YORK (360Dx) – A team led by researchers at Harvard Medical School and the H. Lee Moffitt Cancer Center and Research Institute have identified links between radiological features and molecular and clinical measures in lung cancer.

In a study published last month in the journal eLife, the scientists looked at 351 lung cancer patients across two independent cohorts, developing and validating imaging features that appear predictive of activation status in certain cancer-linked pathways and demonstrating that radiomic data can add prognostic value to more commonly used measures like genetic and clinical information.

Imaging has long been an essential medical technology, with various methods used for diagnosing and assessing patients with a wide variety of conditions. Over the last decade, the field of radiomics has emerged with researchers exploring ways to extract additional data from commonly ordered imaging studies. Somewhat analogously to genomics and DNA or proteomics and proteins, radiomics aims to mine images for hundreds or thousands of discrete features that could prove medically relevant.

"Basically, the idea is to convert images into structured data," said Robert Gillies, director of the experimental imaging program at Moffitt and author on the eLife paper, "because right now images are visualized as pictures, not data."

By considering imaging studies as data, radiomics researchers aim to extract hundreds, even thousands, of features that could inform physicians as to, for instance, a tumor's stage or subtype or likely aggressiveness.

"We can extract [data on] hundreds of different traits of a tumor, for instance, spiking and infiltration and other different aspects, completely automatically," Gillies said. He noted that the field has benefited from breakthroughs in fields like artificial intelligence and pattern recognition where extracting data from images is similarly a key challenge.

"The algorithms we are using are similar to the algorithms that [researchers] use for computer vision, for robots, even for a car driving through the streets," he said. "It's a very similar approach that we apply."

Gillies added that while the field is still working to define specific imaging features and determine which are most relevant to various disease questions, "more and more there is a consensus being reached about feature definitions and how to extract them."

Hugo Aerts, senior author on the eLife study and director of the Computational Imaging and Bioinformatics Laboratory at Harvard and the Dana Farber Cancer Institute, said that the field has, to date, identified around "several hundred features that are really informative."

More recently, he noted, the field has begun working to link these radiomic features to a given patient's underlying biology, including to genomic data. In the eLife study, the researchers analyzed two independent lung cancer cohorts, one of 262 North American patients and the other of 89 European patients. Looking at patient CT scans, they extracted 636 radiomic features, which they then grouped according to their associations with underlying molecular pathways, establishing in this way 13 radiomic-pathway groups.

Investigating these groupings in the context of patient clinical data, the researchers found that three were significantly prognostic for overall survival, 10 were linked to tumor stage, and five were strongly associated with tumor histology.

They also identified links between groupings and molecular pathway activation and mutational status. For instance, the authors noted that the shape features sphericity and compactness contained in one of the groups predicted NF-κB activation mediated by the lung cancer-associated protein TRAF6, while several radiomic-pathway groups were linked to the presence of the known lung cancer driver mutations EGFR and KRAS.

Adding radiomics data to an existing prognostic gene signature (originally described in a 2010 PLOS One paper) and clinical data, the researchers found that a combined radiomic-genetic-clinical model had more prognostic power than either a radiomic-clinical model or a clinical-genetic model. This finding held for several other previously published prognostic gene signatures they tested.

"We could really see that there is complementary information in this radiographic phenotype compared to the genotype that was assessed by gene expression data," Aerts said.

Gillies suggested one factor that could make cancer radiomics complementary to approaches like genomics or conventional pathology is that while these methods often look at only a small portion of the tumor, imaging collects data on the entire tumor in situ.

"With most of these genomic or genetic tests, you're either capturing data from surgery or a biopsy or sometimes even a [fine needle aspirate]," he said. "And that may not be representative of what's going on in the whole tumor. For this reason alone the data types should be complementary."

Aerts and Gillies said, though, that while the field has seen interest from radiology firms and pharmaceutical companies, they have had trouble interesting omics-based researchers, and particularly commercial omics outfits.

"We've reached out to a number of [genomic companies] that have developed some of these tests, but they have not evinced any interest," Gillies said.

One omics firm that has looked into radiomics is proteomics company Integrated Diagnostics, which according to President and Chief Science Officer Paul Kearney, is exploring whether radiomics data could improve the performance of its Xpresys Lung test, which is intended to rule out lung nodules detected on CT scans as likely benign.

Given that the test is meant to help evaluate CT scans, radiomic data is perhaps a natural fit, and Kearney said Indi believes that "integrating radiomics markers with molecular markers has enormous potential to increase the accuracy of lung nodule classification."

He added that the company has completed two prospective studies using radiomic and proteomic markers to evaluate lung nodules. Comprising 1,150 subjects enrolled at more than 45 sites, the studies, Kearney said, found that "the integration of radiomic and proteomic markers far exceeded the performance of radiomic and proteomic markers on their own" while also exceeding the performance of standard lung nodule management guidelines.

He said a study covering this work is in the process of being published.

Gillies and Aerts and their colleagues in the field are also working to integrate radiomics data collection, and, ultimately, usage of that data, into existing radiology workflows. Gillies said that Moffitt is instituting radiomics workstations such that as radiologists do their standard readings, radiomic features are extracted and put into a database.

"We call it the radiology reading room of the future," he said.

One advantage in terms of moving radiomics to the clinic is that the systems for extracting data from radiology studies are already US Food and Drug Administration-cleared. This means clinicians could use this data clinically as part of a "decision support system," Aerts said. "In other words, we can provide a clinician [with an assessment] that says, 'Based on the radiomics, the probability of this outcome is this.' Then it's up to the clinician to decide whether or not to use that information."

Aerts noted that more closed diagnostic systems are also in the works, though these will face stiffer regulatory hurdles. He cited as a potential example a radiomic signature that could, in an automated way, distinguish between benign and malignant lung nodules detected on a CT scan, without employing the radiologists' judgement.

"That's a completely different regulatory path," he said, analogizing it to gene signature tests. "To get that into the clinic will require more time."

In large part, the path the technology takes will depend on the clinical question, Gillies said.

"With some questions, the radiologists can make a pretty good call, and this sort of decision support system can help them with that decision," he said. "In other situations where it is more difficult for a human to make a decision, that is where the real impact from AI in radiology will come from."