---
title: DermDepth
emoji: 🩺
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: Metric-scale 3D from a single dermatology photo
python_version: "3.12"
startup_duration_timeout: 30m
license: cc-by-nc-4.0
tags:
  - depth-estimation
  - 3d-reconstruction
  - dermatology
  - medical-imaging
  - skin-lesion
  - moge
models:
  - hcarrion/DermDepth
  - Ruicheng/moge-2-vitl-normal
datasets:
  - hcarrion/D-Synth
---

# 🩺 DermDepth — Monocular Metric-Scale 3D for Dermatology

Official demo for **[DermDepth: Toward Monocular Metric Scale 3D Reconstruction Models for
Dermatology](https://arxiv.org/abs/2607.13010)** (Carrión & Norouzi, MICCAI 2026).

Dermatology is largely a **measurement** problem — clinicians screen and monitor lesions and wounds
by tracking size, border, elevation and texture over time. Those properties are inherently 3D, yet
point-of-care imaging is almost always a single 2D photograph.

DermDepth recovers **metric-scale** 3D from *one* ordinary photo — no depth sensor, no second view,
no ruler in frame. A 2.1M-parameter scale-and-normal head sits on a frozen
[MoGe-2](https://huggingface.co/Ruicheng/moge-2-vitl-normal) backbone and is trained progressively on
**[D-Synth](https://huggingface.co/datasets/hcarrion/D-Synth)** (synthetic renders with pixel-perfect
depth, normals and intrinsics) and then on real clinical data.

Benchmark results from the paper (held-out test sets) — accuracy on your own photograph may be worse:

| | MoGe-2 baseline | **DermDepth** |
|---|:---:|:---:|
| SKINL2 scale ratio (target 1.0×) | 16.10× | **0.87×** |
| WoundsDB scale ratio | 0.62× | **0.91×** |
| DDI scale ratio | 81.0× | **1.95×** |
| Fitzpatrick scale disparity | 10.90 | **1.02** |

## Two checkpoints, on purpose

| Output | Checkpoint | Why |
|---|---|---|
| Metric depth + 3D mesh | `DermDepth_Synth_SKINL2_WoundsDB_DDI.pt` | The paper's best model (D-Synth → SKINL2 + WoundsDB → DDI pseudo-GT); best metric scale and lowest skin-tone disparity. |
| Surface normals | `DermDepth_Synth_Normals.pt` | The dedicated normal-head model. Real clinical normal ground truth is noisy (SKINL2 plenoptic depth has local planar noise; WoundsDB ToF is sparse and offset from RGB), so D-Synth's rendered normals are the only clean normal supervision. |

These are two complete, separately-trained models that do not share a trunk, so the demo loads both
rather than swapping a head.

## Inference resolution

MoGe-2 converts `resolution_level` into a ViT token budget over `num_tokens_range = [1200, 3600]`.
This demo addresses `num_tokens` directly and keeps every option **inside** that range — "Ultra"
(3600) is the model's genuine maximum, and is the default.

## Examples & credits

The **first three** examples are real clinical photographs from **WoundsDB** (Chronic Wounds
Multimodal Image Database, Silesian University of Technology), used under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) at their native 320×240 — the resolution
the paper evaluates WoundsDB at. They are **held-out** cases (the paper splits WoundsDB by case:
1–30 train, 31+ test), chosen for anatomical variety: `case_45` (leg ulcer), `case_33` (hand),
`case_42` (foot).

> Kręcichwost, M., Czajkowska, J., Wijata, A., Juszczyk, J., Pyciński, B., Biesok, M., Rudzki, M.,
> Majewski, J., Kostecki, J., & Pietka, E. (2021). Chronic wounds multimodal image database.
> *Computerized Medical Imaging and Graphics*, 88, 101844.
> [doi:10.1016/j.compmedimag.2020.101844](https://doi.org/10.1016/j.compmedimag.2020.101844)

The **last three** are **synthetic renders** from
[D-Synth](https://huggingface.co/datasets/hcarrion/D-Synth), released under
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — one per Fitzpatrick group
(I–II, III–IV, V–VI). They are renders, **not** patient photographs, and imply no clinical diagnosis.

No DDI imagery is bundled: Stanford's Research Use Agreement prohibits redistributing any portion
of that dataset.

> ⚠️ **Research demonstration only — not a medical device.** Outputs are not diagnostic and must not
> inform clinical decisions. Predictions on out-of-distribution images can fail silently.

## Links

📄 [Paper](https://arxiv.org/abs/2607.13010) ·
🤗 [Model](https://huggingface.co/hcarrion/DermDepth) ·
📊 [D-Synth](https://huggingface.co/datasets/hcarrion/D-Synth) ·
💻 [Code](https://github.com/hectorcarrion/dermdepth)

```bibtex
@inproceedings{carrion2026dermdepth,
  title     = {DermDepth: Toward Monocular Metric Scale 3D Reconstruction Models for Dermatology},
  author    = {Carri{\'o}n, H{\'e}ctor and Norouzi, Narges},
  booktitle = {Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
  year      = {2026},
  eprint    = {2607.13010},
  archivePrefix = {arXiv}
}
```

## License

Demo code: CC BY-NC 4.0, matching the [DermDepth weights](https://huggingface.co/hcarrion/DermDepth).
The MoGe-2 base weights remain under their original license. Example images are under their
respective licenses, credited above.