Source-linked AI summary

LUTSeg: A Longitudinal Multi-Expert Dataset for Ulcer Tissue Segmentation

Karen Sanchez, Carlos Hinojosa, Albert A. Ávila, Andrea C. Riano-Rojas, Diego H. Romero, Jenny C. Páez, Martina Llinás, Bernard Ghanem

arXiv:2608.25866v1cs.CV

TL;DR

Chronic-ulcer tissue segmentation lacks affordable, longitudinal, multi-tissue annotations, especially for leprosy, despite the clinical value of tissue composition. The paper introduces LUTSeg and TiSage, then reports gains over supervised and semi-supervised baselines across most low-label settings on LUTSeg and DFUTissue.

  • Problem

    Multi-tissue, pixel-level wound datasets are scarce because annotation is costly and subjective, particularly for neglected diseases such as leprosy.

  • Method

    The paper introduces LUTSeg and TiSage, which combines multi-scale semantic priors from a frozen medical vision-language model with a teacher-student framework and pixel-adaptive pseudo-label refinement.

  • Results

    TiSage outperforms UniMatch-V2 in six of seven settings and surpasses supervised approaches across all reported splits, with gains including +4.10 mIoU on LUTSeg 1/8.

  • Takeaways & Limitations

    LUTSeg and TiSage establish an initial benchmark for label-efficient wound-tissue segmentation under annotation scarcity and realistic clinical constraints.

Abstract

from arXiv · show

Quantifying wound tissue composition is essential for monitoring chronic ulcer progression and guiding treatment decisions. However, pixel-level annotations are costly, and multi-tissue wound datasets remain scarce, particularly for neglected diseases such as leprosy. We introduce LUTSeg, a longitudinal chronic ulcer dataset comprising 141 images from 39 patients with wound masks and five tissue categories annotated by five expert clinicians, including a multi-expert gold-standard subset for inter-rater agreement analysis. To establish an initial benchmark for LUTSeg, we further propose TiSage, a semi-supervised tissue segmentation framework that integrates multi-scale semantic priors from a frozen medical vision-language model within a teacher-student architecture. We evaluate TiSage on LUTSeg and DFUTissue, showing improvements over supervised and semi-supervised baselines in most low-label settings. Code & data: https://github.com/carlosh93/TiSage

1 Introduction

LUTSeg addresses the scarcity of longitudinal, pixel-level multi-tissue wound datasets for neglected diseases by introducing a leprosy-related ulcer dataset and a semi-supervised benchmark method. TiSage uses semantic priors from a frozen medical vision-language model to improve tissue segmentation under limited annotation.

  • Existing wound studies commonly focus on boundary segmentation or coarse grading because tissue-level annotations are scarce, while related datasets are private, small-scale, or lack tissue labels.
  • LUTSeg contains longitudinal pixel-level tissue annotations for leprosy-related ulcers, an underrepresented neglected-disease setting.
  • The dataset includes five tissue categories and a 46-image multi-expert gold-standard subset annotated by five clinicians for agreement analysis.
  • TiSage integrates semantic priors from a frozen medical vision-language model into a teacher-student framework for low-label tissue segmentation.

2 LUTSeg Dataset

LUTSeg collects longitudinal ulcer images from 39 patients and provides wound and five-category tissue masks created through a structured multi-expert protocol. Its gold-standard subset enables inter-rater analysis, which shows variable agreement across tissue classes.

  • Data acquisition: 141 images from 39 patients were acquired during routine wound-care visits over 21 months, averaging 3.615 ± 1.695 visits per patient.
  • Annotation protocol: Five specialized clinicians independently annotated wound boundaries and segmented visible wound tissue at pixel resolution.
  • Annotation protocol: The annotation protocol defined Epithelial, Slough, Granulation, Necrotic, and Other tissue categories, allowing Other for visually ambiguous regions.
  • Gold-standard subset: 46 images from 9 patients formed a multi-expert gold-standard subset annotated independently by all five physicians.
  • Inter-rater agreement: ICC agreement was moderate for Necrotic (0.63), Slough (0.55), and Granulation (0.51), lower for Epithelial (0.38), and approximately zero for Other.

3 TiSage Method

TiSage builds semantic priors from frozen MedSigLIP superpixel embeddings at multiple scales, then adaptively fuses them with EMA teacher predictions to refine pseudo-labels for semi-supervised training.

  • Framework overview: TiSage integrates a MedSigLIP-based superpixel prior, multi-scale fusion, and pixel-adaptive teacher–prior fusion into a teacher-student framework.The prior is designed to improve pseudo-label quality in low-label regimes.
  • Semantic prior construction: SLIC superpixels are padded, resized to 448×448, encoded by frozen MedSigLIP, and classified with a frozen linear head.The classifier is trained using majority-voted ground-truth labels and class-balanced cross-entropy.
  • Semantic prior construction: Region-level class probabilities are broadcast across their superpixels to produce a dense per-pixel semantic prior.This converts region predictions into spatially aligned pixel-level guidance.
  • Multi-scale fusion: Coarse and fine superpixel priors are fused in log-probability space to balance spatial smoothness against boundary precision.The fused representation is obtained by applying softmax to the weighted sum of log priors.
  • Pixel-adaptive fusion: Teacher and prior probabilities are fused in log space, giving the prior greater influence when teacher confidence is low and negligible influence when confidence is high.The fusion uses a confidence-dependent weight and produces calibrated pseudo-labels for unlabeled training views.
  • Training objectives: Calibrated pseudo-labels supervise strongly augmented views through hard and entropy-weighted soft objectives, alongside supervised cross-entropy.The overall objective combines labeled and unlabeled losses.

4 Experiments

TiSage is evaluated against supervised and semi-supervised baselines on DFUTissue and LUTSeg under low-label regimes. Results show stronger overall and per-class segmentation, especially for ambiguous or variable tissue categories, while qualitative outputs have more coherent boundaries.

  • Datasets and metrics: TiSage is evaluated on DFUTissue and LUTSeg using standard mIoU and Dice metrics under low-label regimes.The experiments include 1/4, 1/8, and 1/16 labeled settings, with the EMA teacher used at inference unless otherwise stated.
  • Comparison with baselines: TiSage outperforms UniMatch-V2 in six of seven settings and surpasses supervised approaches across all evaluated splits.Its largest reported gains over UniMatch-V2 are +2.42 and +1.60 mIoU on DFUTissue Fixed and 1/4, and +4.10 mIoU on LUTSeg 1/8.
  • Comparison with baselines: At DFUTissue 1/16, TiSage trails UniMatch-V2 by only 0.47 mIoU.This is the one setting identified where TiSage does not outperform UniMatch-V2.
  • Ablation studies: Multi-scale semantic guidance primarily benefits minority and ambiguous tissue classes under annotation scarcity.The MedSigLIP prior alone is insufficient for high-quality segmentation, while multi-scale fusion improves mIoU over single-scale variants.
  • Qualitative results: TiSage produces more coherent boundaries and fewer fragmented predictions than UniMatch-V2 in ambiguous LUTSeg regions.The qualitative comparison covers two random LUTSeg samples and attributes this pattern to pixel-adaptive semantic fusion.

5 Conclusions

The paper introduces LUTSeg for longitudinal, multi-expert tissue segmentation of ulcers caused by a neglected tropical disease and proposes TiSage for annotation-scarce settings. Experiments on LUTSeg and DFUTissue report consistent gains over established baselines, positioning the work as a benchmark for efficient tissue segmentation under realistic clinical constraints.

  • Dataset: LUTSeg is a longitudinal, multi-expert, pixel-level wound tissue segmentation dataset for ulcers caused by a neglected tropical disease.The dataset addresses tissue segmentation in an underrepresented disease setting.
  • Framework: TiSage is a semi-supervised framework that leverages semantic guidance to improve robustness under annotation scarcity.Its benchmark evaluation uses LUTSeg and DFUTissue.
  • Conclusion: Experiments on LUTSeg and DFUTissue show consistent gains over established baselines.The paper presents the combined dataset and framework as a benchmark for efficient tissue segmentation under realistic clinical constraints.
Loading 2608.25866v1…