Multiagent language-model and CNN system scores BT-RADS on glioma follow-up MRI
A system pairing a locally run 20-billion-parameter gpt-oss model, which extracts clinical variables from notes, with a scoring agent applying BT-RADS logic to tumour volumetrics classified 492 post-treatment glioma MRI examinations with 76.0% accuracy against an expert reference, against 57.5% for the initial clinical assessments.
- Model
- gpt-oss 20B (OpenAI, open weights) extractor agent with a convolutional neural network scorer
- Field
- Medicine
- Date
- 2026-10-21
- Human collaborators
- Mohamed Sobhi Jabal, Evan Calabrese
What was found
BT-RADS standardises response assessment after glioma treatment but requires combining imaging trends with medication effects and radiation timing. 509 consecutive examinations from 2017 to 2021 at one centre were evaluated retrospectively, 492 of them evaluable. Agreement with the reference had a Cohen kappa of 0.71. Sensitivity was high in categories that depend on clinical context (100% for BT-1b, 92.7% for BT-1a, 87.5% for BT-3a) and moderate in those that depend on thresholds (57.1% to 74.8%), and positive predictive value for BT-4 was 92.9%.
Novelty check
Europe PMC, 2026-10-29: TITLE_ABS:"BT-RADS" returned 28 records, mostly validations of the scoring system itself; the earlier automation is a 2025 AJNR paper on automated AI-based response assessment by the same Duke group, including Zhang, LaBella, Houk, Rudie, Mazurowski and Calabrese. This paper extends that line with an agent that extracts clinical variables from notes.
Caveats and known objections
Retrospective and single-centre, developed and evaluated by the same group, with no vendor and no funding stated in the abstract. The comparator is the initial clinical assessment rather than a blinded prospective read, the reference standard is expert annotation, and the endpoint is classification accuracy, a surrogate. No external validation. Running the language model locally keeps notes on site. Autonomy is ai-led on the reading this registry applies to retrospective model-output studies: the score the study measured is the system's own, while in the decision-support use the authors propose a clinician would review it.
No independent check is recorded on this entry yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Multiagent language-model and CNN system scores BT-RADS on glioma follow-up MRI. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-10-21-bt-rads-agentic-glioma-scoringBibTeX
@misc{whataifound-dukeuniversitymedicalcenter-2026-scoring,
title = {Multiagent language-model and CNN system scores BT-RADS on glioma follow-up MRI},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Duke University Medical Center. Verification: Peer reviewed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2026-10-21-bt-rads-agentic-glioma-scoring}
}