ksanchez84/LUTSeg

1---2license: cc-by-4.03size_categories:4- n<1K5task_categories:6- image-segmentation7pretty_name: LUTSeg8tags:9- medical10- wound-care11- semantic-segmentation12- longitudinal13- multi-expert14- leprosy15- image16---17 18# LUTSeg19 20**LUTSeg: A Longitudinal Multi-Expert Dataset for Ulcer Tissue Segmentation**21contains pixel-level tissue annotations for longitudinal, leprosy-related chronic22ulcer images.23 24- **Paper:** [LUTSeg: A Longitudinal Multi-Expert Dataset for Ulcer Tissue Segmentation](https://huggingface.co/papers/2608.25866)25- **Code:** [carlosh93/TiSage](https://github.com/carlosh93/TiSage)26 27The accompanying TiSage paper was selected as a **Spotlight** at the Eleventh28ISIC Skin Image Analysis Workshop @ MICCAI 2026.29 30## Dataset Summary31 32- 141 images from 39 pseudonymized patients33- 111 training images and 30 validation images, split at the patient level34- Longitudinal acquisition over 21 months35- Binary wound masks and six-class tissue masks, including background36- A 46-image gold-standard subset from 9 patients annotated by five clinicians37- Per-clinician masks and inter-rater agreement artifacts for the gold subset38 39LUTSeg reorganizes images collected in the SIMATEC project by patient and time40and adds new expert tissue labels. The source images originate from the41[CO2Wounds dataset](https://data.mendeley.com/datasets/nkw5gx57hw/1) described42in the [original study](https://doi.org/10.1016/j.compbiomed.2023.107753).43 44## Labels45 46| ID | Class |47|---:|---|48| 0 | Background |49| 1 | Epithelial tissue |50| 2 | Slough |51| 3 | Granulation tissue |52| 4 | Necrotic tissue |53| 5 | Other |54 55`Masks/` stores single-channel tissue IDs. `Wound_Masks/` stores binary masks56with values 0 and 255. `Masks_RGB/` provides visualizations and must not be used57as training targets.58 59## Repository Layout60 61```text62Images/                         source RGB images63Masks/                          tissue-label masks64Masks_RGB/                      colorized tissue-mask visualizations65Wound_Masks/                    binary wound masks66metadata.jsonl                  paired files and sample metadata67train.txt, val.txt              full-supervision patient-level split68splits/                         full, 1/4, 1/8, and 1/16 paper splits69gold_standard/                  multi-expert masks and agreement artifacts70checksums.sha256                release integrity manifest71```72 73The identifiers in paths and metadata are dataset-internal pseudonyms. They are74not hospital identifiers or patient names. Clinician and reviewer identifiers75are also permanent pseudonyms.76 77## Download and Use with TiSage78 79Download directly into the location expected by TiSage:80 81```python82from huggingface_hub import snapshot_download83 84snapshot_download(85    repo_id="ksanchez84/LUTSeg",86    repo_type="dataset",87    local_dir="data/LUTSeg",88)89```90 91The resulting `data/LUTSeg/Images`, `data/LUTSeg/Masks`, `train.txt`, and92`val.txt` paths work directly with the code at93[carlosh93/TiSage](https://github.com/carlosh93/TiSage).94 95For Hugging Face Datasets, `metadata.jsonl` uses multiple `*_file_name` fields96to pair each image with its tissue, visualization, and wound masks. The97`split` column distinguishes training and validation samples.98 99## Annotation Protocol100 101Five clinicians with wound-care and skin-tissue expertise used a standardized102interface. Wound boundaries were delineated first, followed by pixel-level103annotation of epithelial, slough, granulation, necrotic, and other tissue. For104the 46-image gold subset, all five clinicians annotated every image. A single105reference mask was selected by anonymized clinician voting; ties used a106fixed-seed random selection. The released inter-rater files support the107paper-reported ICC and pairwise Dice analyses.108 109## Ethics and Privacy110 111Acquisition followed the Declaration of Helsinki. All data were anonymized,112written informed consent was obtained from all participants, and the study was113approved by the participating hospitals' ethics committees (Approval Nos.11405-21 and 30-11-25).115 116The release process removes EXIF, GPS, XMP, comments, and editing metadata from117all images. Visual content should still be treated as sensitive medical data118and handled according to applicable institutional and legal requirements.119 120## Intended Uses and Limitations121 122LUTSeg is intended for research in wound tissue segmentation, longitudinal123wound analysis, annotation variability, and label-efficient learning. It is not124a medical device and must not be used alone for diagnosis or treatment.125 126The dataset is small, represents a specific clinical and disease context, uses127smartphone imagery, and contains substantial class imbalance and inter-rater128variability. Performance may not transfer to other populations, institutions,129cameras, wound etiologies, or care settings without additional validation.130 131## License132 133LUTSeg is released under the Creative Commons Attribution 4.0 International134License (CC BY 4.0). Users must provide appropriate attribution and preserve the135dataset citation. This is the same license as the original CO2Wounds source136image release; LUTSeg's new annotations, splits, and metadata are distributed137under the same terms.138 139## Citation140 141Please cite *LUTSeg: A Longitudinal Multi-Expert Dataset for Ulcer Tissue142Segmentation*, Eleventh ISIC Skin Image Analysis Workshop @ MICCAI 2026. The143final proceedings BibTeX will be added when the bibliographic record is public.