Clinically credible MRI radiogenomics in adult-type diffuse gliomas: intended use, transportability, and claim-conditioned evidence

MRI radiogenomics can recover molecularly informative signal from adult diffuse glioma, but the field is now limited less by proof of predictability than by uncertainty about what a prediction is entitled to support. A model that enriches a research cohort, informs a clinician’s preoperative hypothesis, enters patient counseling, prioritizes molecular testing, or influences treatment is making progressively different claims. Those claims cannot be justified by the same validation evidence. This critical narrative review examines MRI radiogenomics in adult-type diffuse glioma through three linked questions: intended use, transportability, and claim-conditioned evidence. We distinguish six transportability domains—taxonomy, population, acquisition, computational pipeline, molecular reference standard, and workflow—and use recent multicenter and cross-dataset studies to show how apparently strong internal performance can degrade after transfer. We also examine a less visible problem: the molecular label itself can shift when classification rules, assays, or cut-offs change. Recent work has begun to incorporate calibration, decision-curve analysis, incomplete-sequence robustness, and prospective external validation, but these developments remain uneven and do not by themselves establish that model-guided care improves decisions. We describe seven evidence states, from biological association to prospective impact and lifecycle monitoring, and map them to the clinical claim being made rather than presenting them as a universal development ladder. The resulting framework places particular weight on contemporary WHO CNS5 alignment, probability calibration in the target population, incremental value over expert imaging and existing molecular workflows, and live clinical evaluation for high-consequence uses. MRI radiogenomics is closest to clinically credible preoperative molecular support for selected endpoints; evidence is substantially weaker for decision-changing or tissue-substituting claims. The relevant question for the next phase is whether a model resolves clinically meaningful uncertainty at the point of decision, not whether another modeling iteration yields a marginal increase in AUC.