TiSage, a method for marking tissue in ulcer images, outperformed UniMatch-V2 in six of seven reported benchmark settings, including tests that used only a fraction of the labeled data. It trailed UniMatch-V2 by 0.47 mIoU in the DFUTissue one-sixteenth split, so the advantage was not universal.
The scores measure how accurately software assigns tissue labels to image pixels. They are image-analysis results, not evidence that an ulcer heals faster or that a treatment works.
A longitudinal ulcer-image benchmark
The work introduces LUTSeg, a dataset of 141 leprosy-related ulcer images from 39 patients, collected during routine wound-care sessions over 21 months.
Five specialized clinicians independently marked tissue at pixel level through a standardized labeling interface. For a reference subset, 46 images from nine patients were selected for greater tissue diversity and multiple follow-up visits, then annotated independently by all five clinicians. Images were split by patient, and ties in the consensus masks were settled by fixed-seed random selection.
That reference subset allows the study to compare clinicians' labels, but most images were not annotated by all five clinicians. The remaining patients were distributed among annotators for separate labeling, and the selected subset may not represent the full cohort.
Clinicians did not agree equally on every tissue
On the 46-image reference subset, agreement in tissue proportions was moderate for Necrotic tissue, with an ICC of 0.63 and a 95% confidence interval of 0.51 to 0.75. The figures were 0.55 for Slough, with an interval of 0.42 to 0.68, and 0.51 for Granulation, with an interval of 0.38 to 0.65. ICC, or intraclass correlation coefficient, is a measure of how closely repeated measurements agree.
Agreement was lower for Epithelial tissue, with an ICC of 0.38, and was approximately zero for Other, with a 95% confidence interval from -0.08 to 0.11. Across 460 pairwise comparisons between annotators' masks, the mean Dice score was 0.814, the median was 0.860 and the minimum was 0.187. Dice measures how much two labeled regions overlap.
The figures show that agreement between annotators varied substantially across tissue categories: some comparisons of annotators' masks were close, while others differed considerably. The authors interpret this pattern as evidence that wound-tissue phenotyping contains a substantial subjective element, especially around ambiguous categories.
Testing performance when labels are scarce
TiSage combines a frozen MedSigLIP-based superpixel prior with multi-scale fusion, pixel-adaptive teacher-prior fusion and entropy-weighted supervision. The study placed it in a semi-supervised teacher-student setup and compared it with supervised DeepLabV3+-R50 and DINOv2-DPT, as well as semi-supervised FixMatch and UniMatch-V2.
The scores were reported as mIoU, or mean intersection over union, along with Dice and F1 measures. Tests used one-quarter, one-eighth and one-sixteenth labeled-data splits within teacher-student training. All methods used the same backbone and training schedule, and results were averaged over random seeds 0, 1 and 2.
The method was also tested on DFUTissue, a public diabetic-foot-ulcer dataset with pixel-level tissue annotations. Its sample size is not reported in the supplied analysis, and its tissue categories differ from LUTSeg's, making direct score comparisons more difficult.
The strongest gains came in selected comparisons
TiSage reached 72.36 mIoU and 83.05 F1 on DFUTissue Fixed. On LUTSeg with one-eighth of the labeled data, it reached 31.70 mIoU and 39.25 Dice. These are benchmark point estimates, not measures of patient recovery.
Compared with UniMatch-V2, the reported mIoU advantage was 2.42 on DFUTissue Fixed, 1.60 on DFUTissue at one-quarter labels and 4.10 on LUTSeg at one-eighth labels. TiSage also surpassed the supervised approaches across all reported splits.
At the one-eighth label level, class-specific IoU gains were 10.3 for Fibrin on DFUTissue, 9.8 for Slough on LUTSeg and 12.7 for Granulation on LUTSeg. These were point estimates from one label regime without uncertainty intervals.
What the benchmark cannot answer
Nothing in the reported endpoints shows that TiSage improves healing, infection detection, treatment selection or another patient outcome. The evidence covers clinician-label agreement and image-segmentation scores on LUTSeg and DFUTissue.
The study also does not show generalization beyond the LUTSeg sample, its acquisition setting or the public DFUTissue benchmark. Whether performance transfers to external sites, devices and ulcer causes, or helps longitudinal monitoring and treatment decisions, remains open.
The two datasets use different tissue taxonomies, and the agreement analysis found lower consistency for Epithelial and approximately zero agreement for Other. Further work would need to address ambiguous boundaries and those categories before broader clinical use could be assessed.
Model results were averaged over seeds 0, 1 and 2, but the supplied analysis reports no model-level confidence intervals or significance tests. That leaves uncertainty about how stable the reported differences would be under repeated experiments.
Component tests point to a mixed picture
In an ablation test, a fused prior-only setup reached 52.11 mIoU on DFUTissue, compared with 48.05 for a fine-only setup. On LUTSeg, the corresponding figures were 26.28 and 25.44. The analysis still judged the prior alone insufficient for high-quality segmentation.
Removing entropy-weighted KL changed DFUTissue one-eighth mIoU by -0.60. On LUTSeg one-eighth, performance varied by no more than 0.35 mIoU as τ ranged from 0.80 to 0.95, with the best result at αmax = 0.25; no statistical uncertainty was reported for these tests.
Publication details
The supplied document is an arXiv preprint version 1, dated 26 Aug 2026. Journal publication and peer-review status are not reported in the supplied metadata. The authors declare no conflict of interest, and the work was supported by the KAUST Center of Excellence for Generative AI under award number 5940.
The abstract provides a code and data repository at https://github.com/carlosh93/TiSage. Further access and licensing details are not reported in the supplied material.
Paper data and sources
Original title: LUTSeg: A Longitudinal Multi-Expert Dataset for Ulcer Tissue Segmentation
Authors: Karen Sanchez, Carlos Hinojosa, Albert A. Ávila et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text