Separating active irradiation from passive outcome learning in radiotherapy: a finite-model computational benchmark
Abstract
Apparent gains from informative irradiation may reflect treatment reoptimization or better use of outcomes from static treatment. We tested whether changing irradiation adds value beyond optimized static treatment, passive outcome learning, and data-dependent stopping. Three synthetic three-dimensional geometries supplied 51 parent fluences. Four known response scenarios generated Bernoulli risks from normalized organ-dose contrasts at uncalibrated slopes of 0, 4, 16, and 64; one unknown scenario applied to all 2,816 hypothetical uses. Loss included acquisition and deployment. We compared fixed passive and locally selected fixed active sampling, passive optimal stopping, and a restricted one-checkpoint, one-switch active policy with a shared terminal menu. Fixed active sampling improved on fixed passive sampling in three of nine nonzero-slope conditions, but never on passive stopping. One-switch adaptation added at most 4.00625e-5 expected model events avoided. With the policies held fixed, two of four increments above the numerical reporting threshold reversed sign under alternative integration; all four became negative at some allowed prior weighting. Adding three generated probes changed none of 72 paired policy values, and none was used on a positive-probability path. The tested probes did not demonstrate stable incremental value beyond optimized passive stopping. Treatment opportunities, stopping permissions, acquisition costs, and input sensitivity must be separated when attributing benefits to active irradiation. These are restricted finite-model results, not evidence about unrestricted active learning or clinical safety and effectiveness.
Related
- OpenGray: a benchmark for evaluating the decisions of radiotherapy planning agents
- Computational Fluid Dynamics and Discrete Dipole Approximation Applied to Predict Airborne Viruses Disinfection by Ultraviolet Irradiation
- DoseRAD2026: A benchmark dataset for AI‐accelerated photon and proton dose calculation in radiotherapy
- Separating charm and bottom decay electrons with point-cloud deep learning on the charged-hadron environment
- MLIP Detective: Active Failure Mode Discovery Beyond Benchmark Scores for Machine-Learning Interatomic Potentials
BibTeX
@misc{wang2026separating,
title = {{Separating active irradiation from passive outcome learning in radiotherapy: a finite-model computational benchmark}},
author = {Yuntao Wang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.14222},
eprint = {2610.14222},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.14222},
}
RIS
TY - UNPB
TI - Separating active irradiation from passive outcome learning in radiotherapy: a finite-model computational benchmark
AU - Yuntao Wang
PY - 2026
DA - 2026/10/09
PB - arXiv
DO - 10.48550/arxiv.2610.14222
UR - https://arxiv.org/abs/2610.14222
AB - Apparent gains from informative irradiation may reflect treatment reoptimization or better use of outcomes from static treatment. We tested whether changing irradiation adds value beyond optimized static treatment, passive outcome learning, and data-dependent stopping. Three synthetic three-dimensional geometries supplied 51 parent fluences. Four known response scenarios generated Bernoulli risks from normalized organ-dose contrasts at uncalibrated slopes of 0, 4, 16, and 64; one unknown scenario applied to all 2,816 hypothetical uses. Loss included acquisition and deployment. We compared fixed passive and locally selected fixed active sampling, passive optimal stopping, and a restricted one-checkpoint, one-switch active policy with a shared terminal menu. Fixed active sampling improved on fixed passive sampling in three of nine nonzero-slope conditions, but never on passive stopping. One-switch adaptation added at most 4.00625e-5 expected model events avoided. With the policies held fixed, two of four increments above the numerical reporting threshold reversed sign under alternative integration; all four became negative at some allowed prior weighting. Adding three generated probes changed none of 72 paired policy values, and none was used on a positive-probability path. The tested probes did not demonstrate stable incremental value beyond optimized passive stopping. Treatment opportunities, stopping permissions, acquisition costs, and input sensitivity must be separated when attributing benefits to active irradiation. These are restricted finite-model results, not evidence about unrestricted active learning or clinical safety and effectiveness.
ER -