Knowledge-based planning, and when to distrust it

Knowledge-based planning fits a model over previous plans — principal component analysis on the geometry, then stepwise regression — and predicts what DVH is achievable for this patient's anatomy. Eclipse's implementation is RapidPlan; RayStation and Pinnacle have equivalents.

What it is genuinely good for. Consistency and quality assurance. A model flags a plan that sits well outside what its training set achieved for similar anatomy, which catches the plan a tired planner stopped optimizing too early. Studies use it exactly this way — as a second opinion on plans a human made.

Why it inherits your habits. The model predicts what your department achieved, not what is achievable. If everyone in the training set stopped when the rectum met constraint, the model predicts stopping there. One study found the model flagged manual planning outliers on mean dose or D1% for at least one structure, while flagging none of the automated ones — the model learned the automation's floor, not the anatomy's.

Outliers matter more than sample size. Models built for wide distribution are explicitly cleaned of outlier plans first, and models trained on 30 well-chosen cases perform usefully. A small clean training set beats a large dirty one.

On a challenge. A knowledge-based model trained on your own department is a strong start and a weak finish. It gets you to the departmental average quickly, and the departmental average is not what wins a leaderboard.