How plans are scored

Every criterion carries a minimum, an ideal and a number of points. At the ideal you get all of them. At the minimum you get none. Between the two it is linear — halfway is half the points.

Worse than the minimum scores zero on that line and never goes negative, so one bad metric costs you its own points and no more. The direction is taken care of: for a coverage goal higher is better, for an organ limit lower is, and the criteria table shows which by the ≥ or ≤ in front of the number.

It follows that chasing an already-ideal metric earns nothing. Points only exist between the minimum and the ideal, so the plan improves fastest wherever you are furthest from the ideal on the heaviest line.

Breaching one does not cost points. The plan keeps whatever it scored, is marked, and leaves the ranking. It stays visible on the board rather than disappearing, because a plan that scored well and broke a cord limit is the most instructive row on it.

Some hard constraints are checked automatically from the dose. Others — an edited contour, an undeliverable plan — are not, and are enforced by review. Both are listed on the case before you start.

A robot that can enter from any direction is answering a different question than a machine sweeping coplanar arcs, and no optimiser setting closes that gap — so plans are ranked against plans delivered the same way: coplanar, non-coplanar, helical, particle or brachytherapy. Splitting by planning system was considered and rejected: the same console drives machines with opposite geometry — Precision plans both CyberKnife and TomoTherapy — so a vendor board still mixes the matchup it was meant to prevent. The planning system stays visible as a column.

The class is measured from your RTPLAN’s couch angles, not chosen by you, which is why the plan file is required with every submission. A C-arm plan with a couch kick is non-coplanar and competes as one; helical is its own family rather than “coplanar with extra steps”, because its longitudinal falloff is its own physics. A plan whose class cannot be read is scored but joins no board.

Resubmitting is free and the number of attempts is shown next to your rank rather than capped. Iteration is the thing being taught, so hiding it would be teaching the wrong lesson — and a limit only advantages whoever guesses well the first time.

You choose which plan is your entry. Everything else stays on your own record and in your write-up, where the chain of what you tried is worth more than the winner alone.

They are versioned. A change creates a new revision with a written reason, every affected plan is rescored against it, and the change is shown on the case with its date. Nobody is quietly re-ranked.

Scores from different revisions are never mixed on one board. If you were scored under an older one, your card says so.

Scoring reads the RTDOSE you upload and the structure set the case ships. Dose–volume histograms are computed per structure, and the criteria are evaluated against those — the same code for a photon plan, a proton plan or an implant, because a histogram does not know what produced it.

The structure a criterion is measured on is named beside it in the criteria table. Public datasets do not follow TG-263, so a goal called “PTV V100%” may be measured on a contour called Prostate — drawing your own PTV instead would produce a plan scored against a contour nobody scored.

The plan summary — beam count, arcs, monitor units — is read separately from the RTPLAN and does not affect the score, only the board assignment above. A plan file that cannot be read still scores on its dose; it just competes on no board until a readable one is submitted.