QianGroup/willie-benchmark

WILLIE Wound Benchmark

Three public wound datasets unified into a single 5-class taxonomy with fixed splits for classification, segmentation and localization.

The benchmark accompanying WILLIE, published at MLHC 2026.

Developed in the Qian Group, University of Houston.

• Models: QianGroup/willie-weights
• Code and notebooks: GitHub repository
• Paper: MLHC 2026 (link to follow)

---

What this is

Three public wound datasets — FUSeg, AZH and Medetec — mapped onto one 5-class taxonomy with fixed, verified splits and a shared evaluation protocol across three tasks. Before this, the three were benchmarked separately with incompatible label schemes, which made cross-dataset comparison impossible.

The contribution is the unification and the splits, not the images.

Class	Label
diabetic	0
pressure	1
surgical	2
venous	3
no_wound	4

3,535 referenced files across classification, detection and segmentation tasks.

---

Contents

data.tar.gz              images from all three source datasets
index/
  data_index.csv         unified index, 2,525 rows, the authoritative source
  splits.json            split definitions
  label_map.json         taxonomy mapping
  azh_clean_trainval.csv AZH cleaned train/val index
  fuseg_det_boxes.csv    FUSeg detection boxes
manifests/
  cls_train.csv          918   classification
  cls_val.csv            648
  cls_test.csv           234
  det_train.csv          810   detection
  det_val.csv            400
  fuseg_test.csv         200   segmentation test
  5fold_splits_v2.pt     cross-validation folds
  data_census.json       per-source counts
verify_data.py           confirms your images are correctly placed

---

Usage

tar -xzf data.tar.gz
python verify_data.py

Expected output:

TOTAL                      3535        0
All referenced files present. The splits will reproduce.

The script reports per-directory counts and names any missing files, so a partial or mis-nested extraction is caught before you train on it.

Expected layout

data/
├── FUSeg/
│   ├── train/images/   610      train/labels/   610
│   ├── val/images/     400      val/labels/     400
│   └── test/images/    200      (no public labels)
├── AZH/
│   ├── train/   BG 75 · diabetic 139 · "no wound" 75 ·
│   │            pressure 100 · surgical 122 · venous 185
│   └── test/    BG 25 · diabetic 46 · "no wound" 25 ·
│                pressure 34 · surgical 42 · venous 62
└── Medetec/
    ├── diabetic/   48        pressure/  170
    └── toes/       34        venous/    133

The AZH class folder no wound contains a space. Keep it.

All manifest paths are relative to the repository root.

---

Source datasets

Dataset	Source	Terms
FUSeg	fusc.grand-challenge.org	Challenge data-use agreement required
AZH	UWM Big Data Lab, AZH Wound and Vascular Center	See source repository
Medetec	medetec.co.uk	See site terms

Each carries its own licence. Obtain them from their original sources and comply with those terms. The archive here is provided for reproducibility of the published splits and does not grant any rights over the underlying images.

---

Reference results

WILLIE test-set results on this benchmark:

Task	Metric	Score
Classification	Accuracy	91.88%
Segmentation	Dice	91.41%
Localization	AP@0.5	96.23%

Headline numbers use the 5-fold ensemble with test-time augmentation. See the paper for the full protocol.

---

Known limitations

Duplicate images across splits. Verified by byte-level comparison:

• One image identical between AZH train and test (train/surgical/10_0.jpg / test/surgical/99_0.jpg). On the 234-image AZH test set, the maximum effect on the reported 91.88% accuracy is 0.43 percentage points (worst case 91.45%).
• Nine image pairs identical between FUSeg train and validation. Validation is used for model selection only and does not enter the reported test Dice.
• Duplicate copies within splits: FUSeg train 10, validation 6, test 7.
• FUSeg train↔test and validation↔test: no duplicates.

Domain coverage. Foot, pressure, venous and surgical wound photographs from a small number of clinical sources. Imaging conditions, camera hardware and skin-tone distribution are not controlled or documented, and per-skin-tone performance has not been measured. Treat cross-population generalisation as untested.

Class balance. no_wound and background classes derive from AZH only. The taxonomy merges source-specific labels; the mapping is in index/label_map.json.

Research use only. Not validated for clinical decision-making.

---

Citation

@inproceedings{willie2026,
  title     = {WILLIE: A Unified Vision-Transformer Framework and Benchmark
               for Wound Classification, Segmentation, and Localization},
  author    = {Maddikunta, Gopi Trinadh and Qian, Peizhu},
  booktitle = {Proceedings of the Machine Learning for Healthcare Conference (MLHC)},
  year      = {2026}
}

Please also cite FUSeg, AZH and Medetec per their own requirements.

---

Developed in the Qian Group, University of Houston. Advisor: Dr. Peizhu Qian. Computation performed on the UH carya cluster.