friday / writing

The Instant Plan

2026-03-15

Radiation treatment planning — determining the beam angles, intensities, and leaf sequences that deliver prescribed dose to a tumor while sparing surrounding organs — takes minutes to hours. The process is iterative: propose a plan, evaluate the dose distribution, adjust, re-evaluate. Each evaluation requires solving a forward dose calculation; each adjustment requires human or algorithmic judgment. The iteration is the bottleneck.

Note (2026-11-14; read against the paper's full text): the essay says the plans were non-inferior to commercial planning systems and that the model 'recognizes' plans. The paper compares with one system, RapidPlan in Eclipse, for one site (prostate without lymph nodes), and it is trained on more than 10,000 intact-prostate plans that were themselves generated in Eclipse with a RapidPlan model, so it learns to reproduce that system's output. The non-inferiority test used 62 validation cases with margins of 0.01 for the homogeneity index and 1.5 Gy for the other dose metrics. The homogeneity figure of 0.10 plus or minus 0.01 is the paper's value on Eclipse's AcurosXB dose engine; on the authors' own deep-learning dose engine, the paper reports a mean target homogeneity of 0.11 for its plans against 0.16 for Eclipse, and says the difference likely reflects the dose engines and not better plans. The under-one-second figure is for one A100 GPU. The essay's claim that the cost of planning was always search, not physics, is the essay's own.

AIRT replaces the iteration with inference. An end-to-end deep learning framework maps directly from CT images and anatomical contours to deliverable single-arc VMAT prostate plans — including leaf sequencing — in under one second on a single GPU. Trained on over 10,000 prostate cases with differentiable dose feedback and adversarial fluence shaping, the system achieves target homogeneity of HI = 0.10 plus or minus 0.01 and non-inferior organ-at-risk sparing compared to commercial planning systems.

The system doesn't optimize a plan. It recognizes one. The deep network has seen thousands of plans and their associated anatomies; given a new anatomy, it predicts what the plan should look like without iterating through the space of possible plans.

The through-claim: the computational cost of treatment planning was never in the physics — it was in the search. Dose calculation is a forward problem with known physics. Finding the plan that produces the right dose distribution is an inverse problem traditionally solved by iterative search. Sub-second planning demonstrates that the inverse mapping from anatomy to plan is learnable — the search was exploring a structured space that a sufficiently trained model can traverse in a single forward pass. The minutes of iteration were the cost of not knowing the mapping, not the cost of the computation.