SAN ANTONIO - Radiation oncology is already putting artificial intelligence (AI) to work by using the technology for automating mundane tasks. But AI's real value will come when it is able to accurately predict patient outcomes, according to a Sunday talk at the American Society for Radiation Oncology (ASTRO) annual meeting.
While much of the attention regarding AI and healthcare has focused on the diagnostic side -- using it to interpret medical images, for example -- radiation oncology is a perfect fit for AI, given the discipline's heavy use of computers for tasks such as radiation therapy planning. And that's just the beginning, as advances in AI eventually make it possible to achieve the holy grail of healthcare -- accurately predicting patient outcomes.
Sunday's talk featured Kristy Brock, PhD, of the University of Texas MD Anderson Cancer Center in Houston, and Andre Dekker, PhD, from Maastricht University Medical Center in the Netherlands. Brock and Dekker outlined the work that's currently underway at the intersection of artificial intelligence and radiation oncology, and they discussed what needs to be done before radiation oncologists witness its full potential.
Brock provided an overview of artificial intelligence and the subsets of AI, machine learning, and deep learning. Writing an AI algorithm is just the beginning, she noted. It must then be validated with thousands, if not millions, of images to improve its performance. And even then, things can go wrong -- an algorithm that MD Anderson developed became flummoxed when it thought images of patients were in the prone position when actually they were supine.
But thanks to growing computer power and larger image datasets for training algorithms, major advances are being made.
Brock offered a number of potential uses for AI in radiation therapy that seem far-fetched now but might not be so far away:
- Take images acquired with one modality and reconstruct them into another modality, such as CT from MRI or MRI from CT, or create contrast images from noncontrast scans.
- Derive tumor history without doing a biopsy, and use that to develop the most appropriate treatment plan.
- Segment and identify toxicity risk to normal tissue on an individualized basis.
- Perform remote treatment planning for developing countries that might not have access to expertise and personnel.
- Perform real-time adaptive change to treatment plans at the therapy unit.
- Predict outcomes in patients over long-term periods.
This last item is key, Brock emphasized. If enough patient data are input into an AI algorithm, could it be used to predict five-year survival, 10-year survival, toxicity risk, and pathological complete response for individual patients? Could it be used to select which patients should receive more advanced forms of radiation-based therapy, such as proton- or carbon-based approaches?
"Artificial intelligence is an enabling technology, not a robot replacing human technology, in radiation oncology," Brock said. "In the future, we will be able to precisely determine the best treatment with the lowest toxicity and understand these risks and benefits for discussion with the patients. And we will be able to translate this optimal plan into the best delivered plan through personalized adaptive real-time radiotherapy."
In his talk, Dekker discussed what he sees as one of the major advances in AI, which is that the algorithms are becoming less complex and easier for clinicians to use, especially through tools like the Inception deep-learning model, he said. This is leading to the rapid development of AI algorithms; he mentioned a project Maastricht did with a major medical imaging OEM, which was able to develop an algorithm for automated autocontouring of medical images, based on data from Maastricht, in just two and a half months -- and the tool works well, Dekker said.
Dekker predicted that within radiation oncology, AI will first be used for simple tasks such as autocontouring, enabling the use of smaller treatment margins, reducing toxicity, and reducing waiting times for patients. These will help make work easier for radiation oncologists and medical physicists, they but won't necessarily transform the discipline.
That will come once AI can correctly predict outcomes for patients. Recent studies have found that physicians perform only slightly better than chance in terms of predicting which patients will be alive over a particular time period (in fact, one study found that nurses actually outperformed doctors).
"Why is this? It probably is because ... we just see too much data," Dekker said. "Humans -- including doctors -- can only take five things into account when making a decision. But we know that for outcomes there are many more factors that are important ... and this, of course, is where computers and AI excel."
As an example of the practical application of AI, Dekker noted that at Maastricht, they have 50,000 patients who get radiation therapy each year but only 1,600 slots for proton therapy. Which patients should be directed to the proton facility? An AI algorithm could develop treatment plans for both photon- and proton-based treatment and, based on toxicity profiles, direct patients to the appropriate option.
But how do we get to a future in which AI is predicting outcomes? One of the biggest challenges is acquiring the data needed to train algorithms, as Brock noted in her talk. Radiation oncology already collects a massive amount of data, but much of this is lost or is unusable outside the home institution -- which limits the ability of the community to use it to train algorithms.
To fix this problem, Maastricht is participating with a network of other cancer institutions around the world in the Community in Oncology for Rapid Learning (CORAL) network. CORAL asks that institutions use a common data structure that will facilitate sharing by making data interoperable.
CORAL participants began contributing data to the network in June, with the goal of having 20,000 studies in the database by September 1. By the deadline, they ended up with 37,000, Dekker said. Another tool, the Distributed Rapid Learning Dashboard, is a CORAL project that enables users to create customized datasets for training AI algorithms.
Ultimately, the time savings generated by AI in the short term will give radiation oncologists and medical physicists the time and resources to help them develop tools that will transform the specialty over the long term -- and make their jobs more interesting, Brock and Dekker agreed.
"It's going to do the things that the machine can do better, but by offloading that and automating that, it's going to enable us to do the more exciting stuff that is what we really need the people for ... driving the science forward, rather than chucking numbers and stuff into a computer," Brock said.
