Spine Endoscopic Atlas-Tissue: multi-class anatomical segmentation in endoscopic spine surgery

SEA-T provides 3,310 expert-annotated endoscopic spine frames from 65 patients, enabling nine-class pixel-level segmentation with Fleiss’ kappa 0.789 across cervical and lumbar anatomy.

• Why it matters: Endoscopic spine surgery lacks a large, multi-center, pixel-annotated dataset with standardized quality control, limiting robust training and evaluation of segmentation models under domain shift.
• What they did: The dataset was annotated in 3D Slicer with multi-stage quality control, split by center, procedure, and anatomy, and evaluated using Dice-based baselines including nnU-Net v2, CaraNet, MedSAM2, and SAM-based MA-SAM.
• The result: Center-dependent domain shift is observed in baseline performance, and the released PNG/RGB overlays plus NRRD/MRML masks and metadata support reproducible model development and testing.

AI summary of the abstract · may contain errors · not medical advice · read the paper.