RFS news highlights resources, issues, and news relevant to in-training members of the ACR. If you have a topic idea or would like to contribute to the blog, please email RFS Secretary Patricia Balthazar, MD.




Radiation Oncology Corner

What does AI mean for radiation oncology and why should you care?

GettyImages 884295504

Lately, there has been a lot of speculation about the potential impact of AI in the field of medicine. Let us explore what AI could mean for the field for radiologists and radiation oncologists.

 AI is a combination of machine learning algorithms, applied to find a solution to a practical problem. It’s hard to find a professional in the life sciences field who hasn’t heard of the acronym or been intrigued by the recent chatter about the promises and perils of AI. At the same time, AI has been around for a long while, ever since early computer scientists in the 1950s started thinking of algorithms that let machines learn from data. These days, whenever we enter a search term into Bing or Google, we invoke a number of AI agents. Whenever your grammar is autocorrected in Microsoft Word, when an Alexa device provides you with a recipe, when a Tesla car navigates an intersection, or when your license plate is scanned while your drive across a toll bridge, there is a high chance that there is an AI component at work in the background.

The core of any AI technology is an algorithm that can be made to reason about new, previously unseen data, based on conclusions that have been made about similar bits of data in the past. These algorithms are usually probabilistic. That is, they give answers with a degree of certainty attached to them and they really shine where the underlying problem can be posed as one that involves a mathematical computation with large number of possible parameters. Image recognition is a perfect example of this. A guess of what is in the image is made with a degree of certainty and is based on using all image pixels to compute a few numbers, which will then give the user the probabilities of the image belonging to one of the several pre-defined image classes.

The field of radiation oncology is uniquely positioned to use the potential power of AI solutions. Over the past several years, there has been increasing promise in the use of AI to advance this field.

It is increasingly important to understand the possible applications of these algorithms in a radiation oncology setting, as well as understand the potential and the limitations of such algorithms. As noted above, AI technologies rely on large amounts of good prior examples and the output is probabilistic. Complex AI algorithms that could propel radiation oncology forward will rely on quality data and clinical reasoning as much as on software and computer science expertise.

Let’s consider one extreme theoretical example of an AI system used in a radiation oncology clinic.
Current technology would allow someone to build a software system that could automatically and accurately contour incoming patient scans; develop an ideal treatment plan; capture daily image-guided radiation therapy (IGRT) images and adjust the plan during the treatment course; and even re-plan if needed based on anatomical changes.
Such a system will also have information about a patient’s chemotherapy regimen and would be able to make adjustments to it based on outcomes of radiation therapy and optimize the dose delivery based on the patient’s response to chemotherapy . While all the technology exists already, such a system would involve at least the development of the following machine learning models and the integration between them:

  • A segmentation service for medical images
  • A treatment planning engine
  • A segmentation service for IGRT images, which would use prior images to enhance the quality of segmentation updates
  • A service which would optimize radiation plan based on tumor response to radiation
  • A service that learns to update chemotherapy based on prior response to radiation
  • A service that learns to update radiation therapy plan based on prior response to chemotherapy

Each one of these would involve data collection and curation, validation and testing in the clinical setting, and an engineering effort to put the software together. Once live, this theoretical system would also require human oversight and input to ensure that the right modules were being executed at the right times and uncertainties were understood and accounted for according to clinical protocol or the clinician’s best judgement. Hence, engineering such a system would require close partnership between the engineers and the members of the treatment team upfront. The deployment of such a system for each institution would require fine-tuning of the machine learning algorithms to cover the specifics of the particular institution.

Now, if one seriously thinks about implementing such a system, one should start asking lots of questions. How much fine-tuning is needed? Is fine-tuning needed at all? Which of these modules would help your institution and which would not? How hard is it to expand this system to cover additional types of disease or additional disease sites? How can you ensure that whatever probabilistic models are deployed for patient care are rigorously validated, ensuring applicability across a range of clinical settings? How will such a system integrate with other IT systems in the clinic? Will this system improve the clinical outcomes significantly enough to justify the costs?

As AI technologies make their way into the field of radiation oncology in the form of various products, these questions will inevitably come up and it would help to keep them in mind as you think of new AI products.

The example above just scratches the surface of how AI might improve the field. There are many workflows where the mundane tasks could be eliminated, the obscure connections may be leveraged, and the consistency of treatment may be improved. It is important to keep in mind that each possible use can require a huge amount of engineering work, and it is crucial to have clinician input and multi-disciplinary collaborations in order to ensure that the focus remains on systems that will bring high quality and value to our clinics and our patients.

Radiation oncology is complex and patient management is increasingly individualized to each patient. Even an incredibly advanced AI algorithm is not going to replace clinician judgement and the need for personalization in patient care. However, AI is uniquely poised to help eliminate many of the more tedious and time-consuming tasks in radiation oncology, freeing up valuable time for physicians to focus on what really matters – our patients.

Ivan Tarapov, senior software engineer, Microsoft, and Meghan Macomber, MD, MS, ACR Resident and Fellow Section radiation oncology representative and chief radiation oncology resident at University of Washington School of Medicine (This email address is being protected from spambots. You need JavaScript enabled to view it.)

Share this content

Submit to FacebookSubmit to Google PlusSubmit to TwitterSubmit to LinkedIn