CJRonald/burn-segmentation-unetpp-resnet50

Model Card: Burn Segmentation UNet++ (ResNet50)

Model Details

• —Developer: Chih-Jung Huang (CGMH Linkou, Plastic & Reconstructive Surgery)
• —Architecture: Deep Supervision U-Net++ (segmentation-models-pytorch.UnetPlusPlus) with ResNet50 encoder
• —Input: 512×512 RGB image
• —Output: 2-class segmentation (background, burn wound) → binary mask after argmax + threshold 0.5
• —Parameters: 61.3 M (inference), 62.9 M (training, includes 4 auxiliary deep-supervision heads)
• —Precision: FP16 for HF Space deployment, FP32 reference checkpoint preserved
• —Framework: PyTorch 2.4 + segmentation-models-pytorch 0.3.4

Training

• —Dataset: CGMH internal burn wound photograph dataset, N=500 images with LabelMe polygon annotations (private, contains PHI, never released)
• —Augmentation: Albumentations resize + ImageNet normalization (no flips/rotation in inference pipeline)
• —Loss: Cross-entropy on main output + weighted sum of 4 auxiliary heads (weights [0.8, 0.6, 0.4, 0.2])
• —Optimizer: Adam (state stripped from inference checkpoint)
• —Best epoch: 14 (after 218 minutes of training)
• —Source experiment: Research/Burn_Segmentation/BurnSegmentation_Training/results/stage1_deepsupervision/

Evaluation

Improvement over prior baseline:

• —IoU +2.2% (0.8364 → 0.8546)
• —DICE +1.5% (0.9076 → 0.9211)

Intended use

• —✅ Research demonstration of deep-learning segmentation on burn imagery
• —✅ Teaching residents about AI-assisted wound assessment
• —✅ Generating preliminary masks for active-learning annotation pipelines
• —❌ NOT for clinical diagnosis or treatment decisions
• —❌ NOT a substitute for Lund-Browder / Rule of Nines TBSA assessment
• —❌ NOT validated outside the CGMH training institution

Limitations

1. 1.Domain shift: Trained on Asian skin tones, hospital photography conditions. Performance on other populations / lighting setups is unverified.
2. 2.Wound heterogeneity: No distinction between burn depth (1st/2nd/3rd degree). Output is wound-vs-not-wound only.
3. 3.Image quality sensitive: Severe blur, extreme angles, heavy occlusion (dressings) reduce accuracy.
4. 4.TBSA scaling is naive: The "TBSA estimate" shown in the demo is image_area_ratio × 10, which is a placeholder heuristic with no anatomical correction. Use only for illustrative purposes.
5. 5.Single-institution training: External validation pending.

Risks and mitigations

Ethical considerations

• —Training data was collected with appropriate institutional oversight at CGMH Linkou.
• —The model was developed for research and educational purposes.
• —No patient-identifiable data is included in the released checkpoint or demo.
• —Public demonstration images (public_demo_images/) are limited to anonymized, publication-released, or synthetic images.

Citation

@misc{huang2026burnseg,
  author = {Huang, Chih-Jung},
  title  = {Burn Wound Segmentation Demo},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/spaces/CJRonald/burn-segmentation-demo}}
}

Contact

• —Email: research@cgmhburncenter.org
• —ORCID / Google Scholar: see https://cgmhburncenter.org