A multimodal AI second read for knee MRI. Twelve abnormalities, one model, structured for the radiologist who signs the report.
miss rate reported for specific knee MRI findings, varying with reader experience.
résumés per open radiologist vacancy in Russia. New modalities (CT, MRI, PET-CT) keep adding reading time.
for a detailed read of one study. A neural network returns its analysis in about 3 seconds.
Most tools bet on chemistry-grade model tricks. We bet on the physics of the data: every study, whatever the scanner or protocol, becomes the same fixed tensor before a model ever sees it.
3 planes × 3 slices, 140 mm crop, 224 px. Series are mapped into consistent slots; a missing-slot mask records absent acquisitions instead of faking them.
A multilingual parser with negation handling turns ~4,400 free-text reports in ~12 languages into training labels. No patient metadata, no manual structured input.
ResNet-18 baseline scaling to DINOv2/v3 ViT-S with per-finding attention pooling. Rank-space ensembling, evaluated by macro ROC-AUC across all 12 findings.
Output arrives as a structured second-read checklist with confidence per finding. The workflow stays the same; the report is still signed by a human.

“A detailed analysis of a single image takes about 90 minutes. At the same time, a neural network will spend 3 seconds analyzing and interpreting one examination.”





Any vendor. The pipeline standardizes series from 19 sites into a fixed 3×3 slot tensor (three planes, three slices) with a 140 mm crop at 224 px. Sequences that are absent are recorded by a missing-slot mask rather than substituted.
We are interviewing radiologists about workflow fit. Twenty minutes, your schedule. Or send a de-identified knee case and we return the 12-finding read.