Research

Funded by the Foundation of the American Society of Neuroradiology to study radiomics of hippocampal subfields in LATE.

First results on hippocampal subfield radiomics in LATE, followed by a postmortem MRI signature of LATE neuropathological change presented as an oral talk.

In the clinic, the same questions come up at Stanford’s Multidisciplinary Memory Conference, and in teaching: talks on imaging and radiomics of the hippocampus at USC and for Stanford’s neuroradiology fellows.

Radiomic signatures from admission CT, in roughly 900 patients split into discovery and validation cohorts, tracked admission severity and three-month outcome more closely than hematoma volume did.

Of 14 shape features, the maximum diameter in the coronal plane predicted three-month outcome better than volume. Beyond roughly 3.5 cm, outcomes were worse.

A combined radiomic and clinical signature predicted hematoma expansion better than visual markers or the BAT score, holding up in an independent validation cohort with an AUC of 0.64.

In large hemispheric infarction, strokes extending into anterior cerebral artery territory independently predicted worse three-month outcome.

Two first-author pieces: a review of how imaging guides spontaneous hemorrhage care, and a case showing how contrast leaking into subdural collections after angiography can mimic a subdural hematoma on CT before clearing within about a day.

Across 259 aneurysms treated with the Pipeline device, occlusion accumulated from about 38% at six months to 78% at a year and 85% at two years. That curve became a proposed follow-up imaging schedule.

About one in four aneurysms stayed incompletely occluded. Age over 70, size of 15 mm or more, and fusiform shape predicted it.

Two flow diverters, Pipeline and FRED, performed comparably in matched internal carotid aneurysms, with 80% complete occlusion for each.

On antiplatelet therapy: a pharmacy-managed protocol compared with one-time platelet testing before Pipeline placement, and a review arguing for genotype- and response-guided P2Y12 therapy.

Language models that draft the impression of a radiology report, and a funded project on automated report generation. Then the human side: whether natural language tools speed dictation and ease burnout.

Opportunistic screening: an AI model that finds coronary artery calcium on chest CTs ordered for other reasons.

A review of AI in brain tumor imaging concluding that the promise is real, but multicenter validation and standardized image pipelines have to come first.

Tailored language models for personalized radiology education, as an invited RSNA talk. For carotid plaque, a synthesis for WNRS: what AI adds is reach, consistency, and measurements no one can make by eye.

Now at Stanford: informatics work on how neuroradiology studies are hung and read, and on evaluating AI tools before clinical use. Nationally, the ASNR AI Committee and its white paper.

Telerehabilitation after stroke: across 34 studies and 1,025 patients, every study reported improvement from baseline, though the methodological quality of the evidence was modest.

A review of neuroprotective strategies in traumatic brain injury, and why so many have failed a heterogeneous disease.