For a given radiotherapy treatment plan type (e.g. prostate, breast, head & neck, etc), explore the use of an ‘unsupervised’ AI training process to determine the properties of a correct treatment plan. Once trained, the system can determine whether a new treatment plan, which it has not previously seen, looks correct and, perhaps give an associated confidence score and feedback on any items that seem incorrect.
A treatment plan has a range of attributes that can be rigorously defined (there is a well-defined structure of treatment plans). We can also generate metrics, like the size of anatomical contours, the mean dose to contours, dose-volume-histograms (DVHs), average MLC leaf gap size, etc. Therefore, we can list all the attributes that any treatment plan can have and list all the metrics that we would normally use to assess a treatment plan.
You will be given a set of anonymsed plans with relevant attributes and metrics, and this may concern a particular plan type: e.g. prostate or head & neck, say. The idea is to learn to recognise the typical attributes of a given plan type. Once this is done, the AI should be able to determine how likely it is that a new given plan is correct or not. In this case, “correct” means:
- attributes such as beam labels, prescribed dose, number of monitor units, number of fractions, number of beams, set-up fields, etc, conform to the typical attributes for the plans that were used in the training; and metrics like DVHs and MLC motions show characteristics that are typical for the given plan type.
We would need to clearly define the structure of a treatment plan and which attributes/variables to include. The structure may well be complicated. For instance, a plan can have several treatment fields (arcs) and each arc has an associated MLC trajectory, dose matrix, monitor units, etc. Therefore, the attributes of a treatment plan can have sub- attributes. More information can be provided by my collaborators from the Radiotherapy Department, Royal Infirmary Aberdeen.
A few notes:
- In principle, every treatment plan created since 2008 is available in Aria. However, since clinical practice has changed over the years, it would be best to restrict ourselves to recent plans and to choose a plan type that is likely to remain relatively stable in the near term.
- We should choose a plan type that has a high number of patients per year, so that there is sufficient data available.