Myth-busting AI autocontouring

25 Mar 2021

Myth Busting: “Autocontouring will take my job?“, Dosimetrist

Some expertly trained dosimetrist’s in Radiation Therapy have expressed that Artificial Intelligence (AI) is taking away a part of their job that requires an expertly trained eye: organ-at-risk contouring. In some clinics, manual contouring is being substituted with autocontouring. Does this inevitably mean that AI will take the valued dosimetrist role? No. It should help dosimetrists concentrate on more complex cases, be more thorough in the quality checks, and eventually diversify and increase their skills. Autocontouring helps save time and provides one source of consistent contours if a solution like DLExpert is used; DLCExpert will provide all dosimetrists in that hospital the same starting point to quality review and be more exacting. Read more about this in the article: AI Autocontouring, Assistant or Adversary.

FAQ: What is the organ risk for autocontouring?

Autocontouring automatically delineates organs at risk, taking away the manual need to draw around these organs. This task is a critical step in preparing a patient for radiation therapy to ensure that healthy organs surrounding the tumor are spared from radiation to improve patient outcomes. Two forms of autocontouring currently exist Atlas-based contours and Deep Learning, a form of AI.

Atlas-based autocontouring uses a combination of previously contoured anonymized patients, generally 20+. A matching algorithm selects the patient or contour most similar to today’s patient before deformable image registration is used to transfer the contours from the Atlas onto the patient. Deep Learning uses hundreds of anonymized past cases (carefully selected to ensure high-quality, consistent input data) to train an Artificial Intelligence, which learns how to predict the same structures onto new data. Deep Learning contouring is a form of Artificial intelligence based on neural networks mimicking the human brain. Today, Atlas technology is being replaced by superior AI-based Deep Learning technologies. Both solutions deliver autocontouring, but Deep Learning technology is increasingly being shown to provide more usable clinical contours.   (See An evaluation of atlas selection methods for atlas-based automatic segmentation in radiotherapy. 2019, Schipaanboord B et al. for more information). The keystone to any good model is well-curated and robust data sets. Mirada now offers an AI library of contours covering the four anatomical regions, head and neck, thorax, prostate and breast. Remember, not all algorithms are equal: rubbish data and training in equals rubbish out. If you’re looking at different vendors, see the data on their application first.

Myth-busting: Artificial intelligence autocontouring models ‘learn and update in the clinic every time they process a new scan Medical devices and solutions have to pass very rigorous standards before clinical use; a technology that learns whilst working in the clinic would not be permitted, as well as technology that is not good enough yet. All the training is done at Mirada by in-house clinically trained dosimetrists. Once a clinic uses a model, Mirada is always in consultation and feedback from customers to improve DLCExpert and configure it to individual hospital preferences.

Myth Busting: Autocontouring is not as good as humans? You tell us! Our evidence suggests that most clinicians cannot tell the difference between manual contours and ones created by Mirada’s DLCExpert auto-contouring solution. This result is good news, as humans build algorithms to help humans save time, be more accurate, consistent, and focus on what’s important – being with patients and being more exacting. FAQ: How are AI autocontouring tools trained? Mirada receives large anonymized clinical data sets of more than a hundred. Inhouse dosimetrists then curate and consistently contour these. An algorithm is then trained to learn from this, forming the basis of Deep Learning models to be used in the AI library structure set and used in the clinics to contour new data. This technology is where the name DLCExpert originated. If you would like to see how your data looks and performs on DLCExpert, book a consultation. FAQ: Will the software suit me immediately when I run cases in my clinic? As the first clinical worldwide AI autocontouring provider, Mirada recognizes that all hospitals and dosimetrists contour differently. This varied approach to contouring styles is reflected in the differing consensus guidelines and ESTRO and ASTRO guidelines. Mirada begins with these consensus guidelines. We encourage hospitals to see how your data looks and performs on DLCExpert, book a consultation. Myth–busting: Artificial Intelligence should surely handle complex cases such as palliative head and neck case?

Deep Learning models are programmed by humans and roughly reflect human capabilities (just faster), so if a regular dosimetrist would struggle to contour a case, it is likely that the Deep Learning models will too. But it is a little bit more complex. What kind of cases are complex & what is complex? Palliative HN with extensive disease, surgical intervention, mouth bites, face full of artefacts, patients on a tilt board, and a tracheostomy all add complexity. The ability of a Deep Learning autocontouring tool to ‘handle’ a complex case depends on the diversity of training data used to develop it. Like humans, where exposure to information or experience is vital to learn, Deep Learning models also require exposure to different scenarios on a scale so that AI can form a generic picture and solution. Therefore, if a type of patient or a specific patient set–up used is not well represented in the training set, the model may be challenged. This is why in house dosimetrist Yasmin McQuinlan of Mirada and Guys Hospital, London, hopes to increase equity and good ethnic diversification into datasets; watch this space!

FAQ: We use specific colors and naming conventions for our structures. Can we easily use those in your models?

The platform that DLC Expert is powered from can be easily configured to match your hospital preference. Mirada’s Clinical experts will help configure this for you. Myth–busting: Use Cloud and it‘s better – some providers have already me that!

Some companies claim that being on the cloud speeds up the processing of cases, yet like working from home, this depends on network speeds. If your network speed is relatively slow (most hospitals can be), the processing time won’t reflect what is advertised. Although it has some advantages, using a cloud feature reduces the need for some hardware and may present other issues and considerations: If on the cloud, where is your data being pushed to? Is a server somewhere remote? How secure is that server? Is there a contingency if, e.g. if there is an earthquake (e.g. as experienced in Christchurch New Zealand) or near a toilet where the pipes burst (a real example I experienced). Having the cases processed on–site gives you more control and oversight over your patient‘s data.