Self-supervised CT foundation model that segments pathology as visual anomalies, pretrained on 30,000 unlabeled 3D volumes with no annotation.
No providers recorded yet. Browse all providers
A supervised segmentation model can only find what its training set labels. In 3D medical imaging that is a hard ceiling: annotated CT datasets cover a handful of pathology classes, so a model trained on them is blind to everything else in the scan. Screener, from IRA Labs with collaborators at the Marchuk Institute of Numerical Mathematics, Lomonosov Moscow State University, and MBZUAI, sidesteps the ceiling by reframing pathology detection as unsupervised visual anomaly segmentation. Pathological tissue is, by definition, rare relative to healthy tissue; a model that has learned what normal anatomy looks like can flag departures from it without ever being told what a lesion is.
The density-based approach to anomaly segmentation is not new — extract dense features, fit a density model to them, score low-density voxels as anomalous. Screener's contribution is to make both halves of that pipeline learned. Feature extraction uses dense self-supervised learning on unlabeled CT rather than features borrowed from a supervised or ImageNet-pretrained encoder, and the conditioning variables that tell the density model where in the body a voxel sits are learned masking-invariant descriptors rather than hand-crafted positional encodings. Trained on more than 30,000 unlabeled CT volumes, the result transfers to held-out cohorts with no fine-tuning at all.
Screener has three components, all built on UNet-like architectures: a descriptor model trained with a dense VICReg objective at 32 feature dimensions, a condition model producing masking-invariant dense features, and a density model implemented as a normalizing flow. A distillation step then folds the whole pipeline into a single UNet. Pretraining draws on three public collections — NLST, AMOS, and AbdomenAtlas — totalling more than 30,000 unlabeled CT volumes used without curation or filtering. Evaluation covers four external datasets and 1,820 scans: LIDC, MIDRC-RICORD-1a, KiTS, and LiTS. In the unsupervised setting Screener reaches voxel-level AUROC of 0.96 on LIDC, 0.87 on MIDRC, 0.90 on KiTS and 0.93 on LiTS, against 0.87 / 0.76 / 0.76 / 0.80 for the strongest baseline, a patched diffusion model. Dice scores are reported but systematically understated, because the evaluation masks label only certain pathologies while the model detects all of them. In a supervised fine-tuning setting using just 25 labeled cases per fold, Screener-initialized nnUNet reaches 0.31 Dice on LIDC against 0.21 from random initialization. Code is public under the MIT license.
The natural use is triage: sweep a stack of chest or abdominal CT and surface the scans and regions that do not look like normal anatomy, whether or not the finding belongs to a category anyone has curated. That covers incidental findings, rare disease, and the long tail of pathology that supervised detectors never see. As a pretrained backbone it also gives radiology groups a way to build task-specific segmenters from a few dozen annotated cases instead of hundreds.
Screener's result — that domain-specific dense self-supervision beats supervised and ImageNet features for medical anomaly detection by a large margin — is a pointed argument about where the value in medical imaging pretraining actually sits. Its limitations follow from the framing: anomaly scores mark what is unusual, not what is clinically significant, and the evaluation cannot cleanly measure precision because the ground-truth masks are incomplete by construction. The preprint has not yet been peer reviewed, and the released repository provides training pipelines rather than downloadable checkpoints.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.