Researchers release InnerLogic-Surg, open surgical vision encoders trained on millions of surgery images
A team describes an open family of surgical image encoders, built by continued pretraining on roughly 2,900 hours of surgical video, reporting top linear-probe results across 20 tasks.
In plain words
The authors built image-encoding models specialised for surgery by further training an existing general-purpose vision model on a large collection of surgical images. An encoder turns raw images into features that other tools can use, so a stronger encoder can lift many surgical analysis tasks at once. The authors report that their largest model scored best on all 20 benchmark tasks they tested.
What was announced
InnerLogic-Surg is a family of open surgical vision encoders ranging from 21M to 300M parameters, described in an arXiv abstract. Rather than training from scratch, the authors continually pretrained a transformer-based image encoder on millions of surgical images curated from 11,965 internet-sourced surgical videos spanning roughly 2,900 hours and more than fifteen procedure types. The recipe inherits DINOv3's pretraining stage to preserve representation ability after surgical adaptation, and adds a learned feature-upsampling module intended to support high-resolution dense prediction at low token cost.
The evaluation covers workflow, instrument and anatomy recognition, segmentation, detection, and monocular depth. Under the same architecture, the authors report that the largest InnerLogic-Surg model achieved the best linear-probe performance in all 20 surgical tasks, and they argue that surgical continual pretraining yields representations that generalize across procedures, anatomical regions, and task types. They position the encoders as an alternative to the procedure-specific representation learning that has traditionally dominated surgical computer vision. The claims come from the paper's abstract, which does not include full results, training details, or independent verification.
What it does not show
- The evidence is limited to the paper's abstract; full results, methods, and peer review are not available in the supplied passages.
- Benchmark outcomes are author-reported and have not been independently confirmed.
- The abstract does not state model availability, licensing, or release details beyond describing the family as open.
For security and IT teams
- Scope
- The work concerns vision encoders for surgical image analysis, covering tasks such as recognition, segmentation, detection, and depth estimation, not clinical decision-making.
- Status
- The paper appears as an arXiv preprint with cross-listing announce type; no peer-reviewed publication or deployment status is stated in the supplied abstract.
Sources
Source-support review completed 14 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 14 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.