Stereotactic radiosurgery plan optimization with a human-in-the-loop reasoning large language model agent
Abstract
Large language model agents can adjust optimization parameters in a treatment planning system under natural-language instruction, but no study has tested a model post-trained for extended reasoning against a general-purpose model in stereotactic radiosurgery. We compared a reasoning-optimized model (QwQ-32B) with a general-purpose model (Llama 3.1-70B) within SAGE, a locally hosted agent that adjusts optimization-objective priorities in Eclipse, which performs all optimization and dose calculation. Each configuration optimized 41 retrospective single-target brain metastasis cases (18 Gy, single fraction) with fixed clinical beam geometry; plans failing the clinical conformity standard (all but two) then received one standardized physicist instruction. The reasoning-optimized configuration produced a lower conformity ratio (median difference −0.20, 95% CI −0.48 to −0.10; q<0.001 ; lower in 36 of 41 cases) and lower normal-brain V12Gy ( q<0.001 ), improved conformity more after the instruction (difference in change −0.28, p<0.001 ), exceeded 21.6 Gy maximum dose less often (5 versus 12 plans), and produced fewer unparseable outputs (25 versus 122). Against clinical plans it showed no significant difference in PTV coverage, maximum dose, conformity ratio, or gradient index, and its right-cochlear maximum dose was lower, a clinically negligible difference.
Related
- Comparative Analysis of Commercial Independent Dose Calculation Systems for Gamma Knife Stereotactic Radiosurgery
- Benchmarking intermediate dose spillage in single target stereotactic radiosurgery for brain metastases.
- Geometric and dosimetric effects of an automated clustering and plan optimization pipeline for multiple brain metastasis radiosurgery.
- Separating perception from reasoning in vision-language models: a model-free render ceiling for crystal structures
- Toward a social psychology of AI: language-model agents reproduce human-like minimal-group bias
BibTeX
@article{nusrat2026stereotactic,
title = {{Stereotactic radiosurgery plan optimization with a human-in-the-loop reasoning large language model agent}},
author = {Humza Nusrat and Luke Francisco and Bing Luo and Hassan Bagher-Ebadian and Joshua Kim and Karen Chin-Snyder and Salim Siddiqui and Mira Shah and Eric Mellon and Mohammad Ghassemi and Anthony Doemer and Benjamin Movsas and Kundan Thind},
journal = {Scientific Reports},
year = {2026},
month = sep,
doi = {10.1038/s41598-026-73885-x},
url = {https://www.nature.com/articles/s41598-026-73885-x},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Stereotactic radiosurgery plan optimization with a human-in-the-loop reasoning large language model agent
AU - Humza Nusrat
AU - Luke Francisco
AU - Bing Luo
AU - Hassan Bagher-Ebadian
AU - Joshua Kim
AU - Karen Chin-Snyder
AU - Salim Siddiqui
AU - Mira Shah
AU - Eric Mellon
AU - Mohammad Ghassemi
AU - Anthony Doemer
AU - Benjamin Movsas
AU - Kundan Thind
PY - 2026
DA - 2026/09/30
JO - Scientific Reports
DO - 10.1038/s41598-026-73885-x
UR - https://www.nature.com/articles/s41598-026-73885-x
AB - Large language model agents can adjust optimization parameters in a treatment planning system under natural-language instruction, but no study has tested a model post-trained for extended reasoning against a general-purpose model in stereotactic radiosurgery. We compared a reasoning-optimized model (QwQ-32B) with a general-purpose model (Llama 3.1-70B) within SAGE, a locally hosted agent that adjusts optimization-objective priorities in Eclipse, which performs all optimization and dose calculation. Each configuration optimized 41 retrospective single-target brain metastasis cases (18 Gy, single fraction) with fixed clinical beam geometry; plans failing the clinical conformity standard (all but two) then received one standardized physicist instruction. The reasoning-optimized configuration produced a lower conformity ratio (median difference −0.20, 95% CI −0.48 to −0.10; q<0.001 ; lower in 36 of 41 cases) and lower normal-brain V12Gy ( q<0.001 ), improved conformity more after the instruction (difference in change −0.28, p<0.001 ), exceeded 21.6 Gy maximum dose less often (5 versus 12 plans), and produced fewer unparseable outputs (25 versus 122). Against clinical plans it showed no significant difference in PTV coverage, maximum dose, conformity ratio, or gradient index, and its right-cochlear maximum dose was lower, a clinically negligible difference.
SN - 2045-2322
ER -