Table 1.

Summary of current ML research focus in the radiotherapy pathway

Clinical application	Clinical Need	Current ML focus	Well-defined problem?	Well-defined ground truth?	Quantitative measure of correctness?
CT simulation	Image reconstruction quality / dose reduction	Image reconstruction quality / dose reduction	Yes	No	No
MRI simulation	Pseudo CT creation	Pseudo CT creation	Yes	No	Yes
Image fusion	Estimate spatial uncertainty Accommodation of anatomical changes	Registration efficiency Appropriate similarity metric	No - Depends on use-case	No	No
Contouring	OAR/Target Contouring efficiency OAR/Target consistency Target contouring accuracy	OAR/Target Contouring efficiency OAR/Target consistency	Yes	No – Subjective clinical contours used	Yes
Treatment planning	Planning efficiency Plan consistency Determining the plan to deliver the best clinical outcome	Planning efficiency Plan consistency	No – Depends on clinical satisfaction criteria	No – Subjective treatment plans used	No
QA	Efficiency and automation Identification of clinically meaningful errors	Efficiency and automation	n/a	n/a	n/a
Delivery	Dose accuracy in the presence of motion (see Image fusion, Contouring, and Treatment planning) Determining who will most benefit from replanning	Dose accuracy in the presence of motion (see Image fusion, Contouring, and Treatment planning)	No	No	No

ML,machine learning; OAR, organ at risk; QA, quality assurance.

Training machine learning requires a well-defined problem, with a well-defined ground truth, and a simple measure with which to assess effectiveness. The application to QA is not considered in detail, as the status depends on what is being assured and to what degree.