Decision paths help models follow radiology guidelines
Radiology Guideline-derived Decision Graph for Efficient Clinical Decision Support · arXiv:2610.40327
In plain words
Radiology guidelines can spread decision rules across text, tables and definitions. This paper introduces a way to turn those rules into paths through a decision graph. A model answers small questions about a case, and computer rules follow the matching path to a category or management decision.
Across 347 cases from three radiology guideline tasks, guiding Jev with these paths raised overall case accuracy from 57.3% to 90.8%. That average describes the evaluated cases, not every case or setting.
The research
Radiology guidelines distribute decision criteria across prose, tables, definitions and exceptions. When criteria depend on one another, how can a model follow those links reliably enough to assign the guideline's risk category or management recommendation?
GDGraph compiles radiology guidelines into decision paths, using local model judgments and explicit rules to reach traceable case decisions.
The question is whether a model can follow guideline criteria that are spread across sections and depend on one another, rather than miss a condition on the way to a case decision.
GDGraph compiles those criteria into decision paths. In 347 cases across three radiology tasks, Jev's overall case accuracy rose from 57.3% with direct inference to 90.8% with graph guidance. This supports an improvement across the evaluated cases; GPT-4o comparisons also varied by task.
These results come from a retrospective evaluation of curated, text-based cases for three guideline tasks.
The error analysis also found terminology confusions in 11 thyroid cases and difficulties mapping qualitative blood-flow descriptions in three adnexal-mass cases; performance beyond this evaluation is not established.
What the evidence shows
- The guideline-following problem
Radiology guidelines often distribute criteria across prose, tables, definitions and exceptions. Because these criteria depend on one another, a model can depart from the guideline when it assigns a risk category or management recommendation.
Guideline-derived Decision Graph (GDGraph) is a framework that compiles these rules into paths for evaluating cases.
- Turn rules into paths
GDGraph separates guideline construction from case evaluation. A language model extracts local branches and links definitions while compiling the guideline.
During evaluation, a model answers questions about individual criteria, and program rules follow the matching branches to a decision. The resulting path records how the decision was reached.
- Check against source text
The authors first checked whether graph questions, definitions and transition rules were supported by the source guidelines. Manual review found support for 489 of 490 graph questions and definitions and 266 of 276 rules across the three guidelines. This checks graph fidelity to the source, rather than correctness on cases.
- Accuracy depends on comparison
In 347 cases across renal, adnexal-mass and thyroid tasks, Jev's overall case accuracy rose from 57.3% with direct inference to 90.8% with graph guidance.
GPT-4o's graph-guided accuracy was 82.7% overall, versus 80.1% with retrieved guideline passages, but results differed by task: retrieval did better on O-RADS, while graph guidance did better on K-TIRADS.
Deep
- Original paper authors
- Xuewen Zhou, Zhen Chen
- Submitted to arXiv
- License
- arXiv non-exclusive