Artificial Intelligence in Radiotherapy 101

We are here to explore topics around AI in radiotherapy, share practical tips, and address questions as well as biases we with our partners come across in our daily work. Welcome to AI in RT 101 – first up some myths.

The potential of AI in radiotherapy has been acknowledged for a while. Studies around the world have agreed on the benefits of tools powered by machine learning, particularly in mundane tasks. Some of these studies are also listed in our latest reading package. Yet, there are still myths related to AI which in the worst case prevent cancer patients from getting the treatment they deserve. 

“I have seen so many tools and I don’t know if it’s that efficient.”

There are indeed several different options out there. As in any field, the presence of different tools and competition generates pressure for better ones and creates a positive loop of development.  Quite basic. In terms of radiotherapy, it is of course also strongly linked to research. As mentioned in the latest report about AI in Finland, there is a long tradition in AI research in Finland, which has led to cutting-edge technology. While large companies have built their own AI development units, the driving force is the hundreds of innovative startups with roots in universities and, most importantly, research.

So what can we make of this? Let’s take auto-segmentation as an example, The basic idea is that an AI-powered solution makes contours on CT or MR images. An oncologist then checks, approves, or fixes the contours if needed. Since the core idea is to automate a manual task, a proper tool should lead to increased efficiency. However, there are indeed variations in the delivered value of different tools. At MVision, we focus on excellence – both in terms of the tool itself and provided service. By investing in the quality of the algorithms, alignment with standards, effortless implementation and support, it is not only efficient but also improves the quality of the work.

“But it is part of the core skills and professional radar of clinicians to know how to contour.”

There is no question about it. Contouring, or delineation, is a skill any expert participating in treatment planning needs to master. When all the fuzz is removed, a proper AI-powered segmentation tool is merely a way to optimize the use of this skill: Clinicians harness their eye in checking the work of the software instead of drawing every line from scratch. In other words, clinicians still need to maintain their skill of delineation to review and approve the works of the machine. Seamlessly, for the better of the patient.

“Time-saving is not enough, we need quality.”

A state-of-the-art tool also improves the quality of the contours. As we tend to say, our AI-powered segmentation tool that is strongly based on the official consensus guidelines is as if one clinician harnesses the skills of a hundred colleagues. Research by Brouwer et al (2015) shows this in practice, particularly so in one image of an axial CT slice.

Delineation results, Brouwer et al., 2015

CT-based delineation of organs at risk in the head and neck region: DAHANCA, EORTC, GORTEC, HKNPCSG, NCIC CTG, NCRI, NRG Oncology and TROG consensus guidelines by Brouwer et al. (2015)

The image shows delineation results of seven clinicians who took part in the research panel for the parotid glands, spinal cord, pharyngeal constrictor muscles, and the oral cavity. As we can see, the lines clearly differ from each other. While seven clinicians are not even close to a hundred, it raises an important question: If the differences already with such a sample are this evident, how big would they be with a larger one? Undoubtedly even more significant. Brouwer et al. (2015) conclude that incorporating consensus guidelines on a large scale is the solution. 

How does this translate to practice? Our peer-reviewed paper published late last year shows how inter-observer consistency increases with the auto-segmentation method. For pelvic lymph nodes, the Dice score agreement between oncologists is 0.74 and 0.91 for manual method and editing the AI auto-segmentations, respectively. 

Inter-observer consistency

MC – Axial (left) and AEC – Axial (right) from Deep Learning-Based Automated CT Segmentation of Prostate Cancer Anatomy for Radiation Therapy Planning-A Retrospective Multicenter Study by Kiljunen et al. (2020)

In other words, the outcome of the contouring is improved. When combining a state-of-the-art solution with the eye of an expert clinician, it is as simple as that.

Do you have a question about AI in radiotherapy? We would love to hear and give you an answer! Send your question to pr@mvision.ai.


Read more:

Official consensus guidelines used by MVision AI

Brouwer, C. L., Steenbakkers, R. J., Bourhis, J., Budach, W., Grau, C., Grégoire, V., … & Langendijk, J. A. (2015). CT-based delineation of organs at risk in the head and neck region: DAHANCA, EORTC, GORTEC, HKNPCSG, NCIC CTG, NCRI, NRG Oncology and TROG consensus guidelines. Radiotherapy and Oncology, 117(1), 83-90. Available at https://www.eortc.org/app/uploads/2018/02/Atlas-HN.pdf

Kiljunen, T., Akram, S., Niemelä, J., Löyttyniemi, E., Seppälä, J., Heikkilä, J., … & Keyriläinen, J. (2020). A Deep Learning-Based Automated CT Segmentation of Prostate Cancer Anatomy for Radiation Therapy Planning-A Retrospective Multicenter Study. Diagnostics, 10(11), 959. Available at https://www.mdpi.com/2075-4418/10/11/959#cite

Previous Post
AI in RT – Reading Package
Next Post
MVision AI obtains a new CE mark: Ultimate segmentation service now with lymph nodes
Sign up for newsletter

    Menu