Trained checkpoints associated with the paper settings for TiSage: Tissue Segmentation with Multi-Scale Semantic Guidance, released with the code and reproducibility material at carlosh93/TiSage. The paper was selected as a Spotlight at the Eleventh ISIC Skin Image Analysis Workshop @ MICCAI 2026.

Each file contains model, model_ema, optimizer state, epoch, and validation selection metadata. The files are best-student snapshots and contain the EMA state from the same epoch. The paper's Table 1 numbers are selected best-EMA values across training and remain traceable to the code repository's committed logs, so an individual released snapshot can differ from the Table 1 value. The LUTSeg pair is also used by the Figure 4 reproduction script.

Download every released checkpoint with:

Verify the files with segmentation_checkpoints.sha256 from this model repository or the TiSage code repository.

Clone the TiSage repository, install its requirements, download LUTSeg, and run:

The public LUTSeg release has a patient-disjoint validation set and no separate public test split. The evaluator reports per-class IoU and Dice plus their means on the 30-image validation set and requires one CUDA GPU. Append --check-only to validate checkpoint compatibility without inference. The evaluator defaults to the selected student state; pass --state model_ema for the contemporaneous EMA state.

The segmentation network is a DINOv2-Base DPT model. TiSage trains it in an EMA teacher-student framework using frozen MedSigLIP semantic guidance and the small dataset-specific prior heads committed in the code repository. MedSigLIP parameters are not included in these checkpoint files.

Exact configurations, selected seeds, launch commands, result evidence, and the bounded training check are maintained in the TiSage repository. LUTSeg is available at ksanchez84/LUTSeg.

These checkpoints support research reproduction and non-clinical exploration of wound-tissue segmentation. They are not medical devices and must not be used alone for diagnosis or treatment. Results may not transfer to other patient populations, institutions, cameras, wound etiologies, or acquisition settings.

The checkpoint release is distributed under Apache License 2.0 because it contains DINOv2-derived backbone parameters. TiSage code remains MIT licensed, and datasets retain their own terms. Training used the gated google/medsiglip-448 model as a frozen prior; its parameters are not redistributed here. See NOTICE.md for the complete attribution and scope.

Please cite LUTSeg: A Longitudinal Multi-Expert Dataset for Ulcer Tissue Segmentation, Eleventh ISIC Skin Image Analysis Workshop @ MICCAI 2026. Final proceedings metadata will be added when the bibliographic record is public.