ReMIND reads multi-sequence brain MRI the way a neuroradiologist would — reasoning across T1, T2, FLAIR, DWI and beyond to draft structured, radiology-style reports.
Approximate geographic reach from signed-in app visits. Locations are aggregated and used only to show adoption of our research tool.
ReMIND was built around how neuroradiologists actually read — synthesizing complementary contrasts into a coherent narrative, then producing structured language a clinician or a researcher can edit and trust.
DICOM and NIfTI in. T1, T2, FLAIR, DWI, SWI, post-contrast, MRA/MRV — any combination, any protocol.
A vision-language model jointly attends across sequences, surfacing focal and diffuse findings the way an integrated read would.
A structured radiology-style draft you can edit, plus a chat to interrogate findings, sequences and differentials.
Frontier-grade behavior on neuroimaging without frontier-scale compute. Every component is tuned to how brain MRI is acquired, read and reported.
Joint understanding across T1, T2, FLAIR, DWI/ADC, SWI/T2*, post-contrast, and angiographic series — without explicit cueing.
cross-contrastDrafts mapped to the canonical seven sections — Examination, Technique, Comparison, History, Indication, Findings, Impression.
clinician-readyAsk follow-ups across various question types — from yes/no verification to free-form clinical reasoning grounded in the scan.
grounded VQAReMIND was developed on a large collection of clinical cases and held out on independent ones — across academic, community, and international settings — so generalization isn't measured on the data it was trained on.
In high-stakes medical imaging, specialization outperforms general-purpose scale. Domain-specific tuning and rigorous external validation are how AI earns the right to assist a clinician — and that's the bar ReMIND was built to meet.
Architecture, data pipeline, decoding strategy and metrics are published. We welcome scrutiny — that's what peer review is for.
Every output is labeled a draft. The model proposes; the clinician decides. ReMIND assists, it doesn't replace.
Validated across academic, community and international cohorts so performance isn't only a story about one population.
Studies are de-identified and handled with strict safeguards. No data is stored on our servers after the user session ends.
The demo runs on de-identified studies. No friction — just a draft you can interrogate, and a clear boundary about what it is and isn't.