Dermatological practice routinely involves measuring and tracking lesion size, morphology and texture, as critical components of wound or skin cancer screening, monitoring and diagnosis. To accomplish this task, practitioners often image the skin surface with commonly available off-the-shelf camera sensors. This has led to an overwhelming research focus on 2D methods while these objectives naturally benefit from 3D information. In this paper, we demonstrate that dense monocular 3D reconstructions, metric scale measurements and rich surface normal texture estimates are achievable for both dermoscopic and macroscopic cases without the need for additional hardware or multiple captures. We present DermDepth, the first single-view metric scale 3D model for the dermatological domain and D-Synth, the first synthetic dermoscopic dataset with pixel-perfect 3D information. Our experiments show training DermDepth on D-Synth corrects metric scale error from over 16× to under 1.1× for real dermoscopic data, while preserving geometric quality and increasing texture richness. Fine-tuning on a small amount of real clinical samples generalizes our method across three real-world benchmarks spanning the few mm to hundred cm range, diverse skin-tones, chronic wound cases and produces measurements broadly consistent with disease size reported in medical literature. All code, data and models are available at https://github.com/hectorcarrion/dermdepth.
The MICCAI 2026 camera-ready PDF is hosted directly in this repository: Paper-5594.pdf. The arXiv version is embargoed by Springer until approximately September 2027 (one year after open-access publication at MICCAI 2026). The version-of-record will appear in the Lecture Notes in Computer Science proceedings volume shortly before the conference (Sept 27 – Oct 1, 2026, Strasbourg).
DermDepth (★) lies closest to the (1×, 1×) perfect-scale target on both SKINL2 (dermoscopic) and WoundsDB (macroscopic) benchmarks while all foundation baselines over- or under-predict by 2–16×. On DDI, predicted lesion volumes match the ranges reported in dermatological literature (right).
| Method | Params | SKINL2 Scale | SI-d1 (%) | WoundsDB Scale | SI-d1 (%) | DDI Ratio |
|---|---|---|---|---|---|---|
| Depth Anything v3 | 1.4B | 4.16x | 99.6 | 0.67x | 89.1 | 53.6x |
| MapAnything | 1.1B | 10.88x | 99.3 | 0.75x | 80.7 | 156.1x |
| PPD | 831M | 16.21x | 92.0 | 0.66x | 91.3 | 90.4x |
| MoGe-2 | 331M | 16.10x | 100.0 | 0.62x | 91.1 | 81.0x |
| DermDepth_S | 2.1M ft | 1.11x | 100.0 | 0.28x | 91.1 | 9.2x |
| DermDepth | 2.1M ft | 0.87x | 100.0 | 0.91x | 92.6 | 1.95x |
Scale ratio and DDI ratio target = 1.0x. DermDepth_S: synthetic data only. Eval on held-out test sets.
Rows: SKINL2 (dermoscopic), WoundsDB (chronic wounds), DDI (diverse skin tones). Columns from left: input RGB, ground-truth depth (when available), DepthAnything v3, MapAnything, PPD, DermDepth, MoGe-2 normals, DermDepth normals. DermDepth's per-pixel Scale factors hover around 1.0× across modalities while baselines mis-scale by 2–15×.
D-Synth provides 3,132+ rendered dermoscopic samples with pixel-perfect metric depth, surface normals, and camera intrinsics — enabling supervised training where real metric-scale 3D dermatological data does not exist.
Pre-generated D-Synth training data is hosted on Hugging Face: hcarrion/D-Synth.
# CLI download
hf download hcarrion/D-Synth --repo-type dataset --local-dir data/dermdepth_train/dsynth
# Or via Python
python -c "from huggingface_hub import snapshot_download; snapshot_download('hcarrion/D-Synth', repo_type='dataset', local_dir='data/dermdepth_train/dsynth')"To generate your own synthetic dermoscopic training data with metric-scale 3D ground truth:
# Local generation (requires Mitsuba 3 + S-SYNTH assets)
python code/data_generation/generate_dermdepth_dataset.py --num_samples 3000 --output data/dermdepth_train/dsynth
# Or use the Colab notebook for cloud generation
# See notebooks/generate_dermdepth_colab.ipynbDownload evaluation datasets from their original sources:
- SKINL2: Skin Lesion Light Field Dataset (de Faria et al., 2019)
- WoundsDB: Chronic Wound Database (Juszczyk et al., 2020)
- DDI: Diverse Dermatology Images (Daneshjou et al., 2022)
- sDDI: DDI segmentation masks used for ruler annotation
Then prepare for training/evaluation:
# Convert evaluation datasets to MoGe format
python code/data_generation/convert_eval_to_moge.py --dataset skinl2 --input data/SKINL2 --output data/dermdepth_train/skinl2_moge
python code/data_generation/convert_eval_to_moge.py --dataset woundsdb --input data/DB_ALL --output data/dermdepth_train/woundsdb_moge
# Create DDI pseudo-GT from ruler annotations
python code/data_generation/create_ddi_training_data.py --input data/DDI --output data/dermdepth_train/ddi_mogedermdepth/
├── code/
│ ├── analysis/ # Dataset exploration and baseline analysis
│ ├── annotation/ # DDI ruler annotation tools
│ ├── data_generation/ # D-Synth rendering and MoGe format conversion
│ │ ├── generate_dermdepth_dataset.py # Main D-Synth generation script
│ │ ├── convert_to_moge.py # S-SYNTH → MoGe format converter
│ │ ├── convert_eval_to_moge.py # WoundsDB/SKINL2 → MoGe format
│ │ ├── create_ddi_training_data.py # DDI pseudo-GT from ruler areas
│ │ └── depth_utils.py # Depth encoding and intrinsics utils
│ ├── evaluation/ # Metric evaluation scripts
│ │ ├── eval_depth.py # Main depth evaluation (scale, AbsRel, SI-d1)
│ │ ├── eval_ddi_rulers.py # DDI ruler-based evaluation + fairness
│ │ ├── eval_normals.py # Surface normal evaluation
│ │ └── eval_baselines.py # Baseline model evaluation
│ └── visualization/ # Paper figure generation
├── configs/ # MoGe-2 training configs for all experiments
├── notebooks/
│ └── generate_dermdepth_colab.ipynb # Colab notebook for D-Synth generation
├── scripts/ # Evaluation shell scripts
└── LICENSE
DermDepth builds on MoGe-2. Clone and install it first:
git clone https://github.com/microsoft/MoGe.git
cd MoGe && pip install -e .Then install additional dependencies:
pip install accelerate mlflow-skinnyFor D-Synth data generation, you also need Mitsuba 3 and S-SYNTH assets.
Download the MoGe-2 pretrained weights from HuggingFace:
cd MoGe && python -c "from huggingface_hub import hf_hub_download; hf_hub_download('Ruicheng/moge-2-vitl-normal', 'pretrained_moge2.pt', local_dir='.')"Checkpoints are hosted on Hugging Face: hcarrion/DermDepth.
| Filename | Training data | Notes |
|---|---|---|
DermDepth_Synth.pt |
D-Synth only | "DermDepth_S" in the paper (synthetic-only baseline). |
DermDepth_Synth_SKINL2_WoundsDB.pt |
D-Synth → SKINL2 + WoundsDB | Intermediate stage. |
DermDepth_Synth_SKINL2_WoundsDB_DDI.pt |
D-Synth → SKINL2 + WoundsDB → DDI pseudo-GT | Best model (paper "DermDepth" results). |
DermDepth_Synth_Normals.pt |
D-Synth, normal-head emphasis | Alternative for normal-map use. |
# Single checkpoint
python -c "from huggingface_hub import hf_hub_download; hf_hub_download('hcarrion/DermDepth', 'DermDepth_Synth_SKINL2_WoundsDB_DDI.pt', local_dir='checkpoints')"
# All checkpoints
hf download hcarrion/DermDepth --local-dir checkpointsDermDepth uses progressive training with scale-head-only fine-tuning:
cd MoGe
# Stage 1: Synthetic-only (scale head, ~2.5h on 1x A100)
python -m moge.train --config ../configs/dermdepth_exp_a.json --workspace ../output/training/exp_a
# Stage 2: + Real data (WoundsDB + SKINL2)
python -m moge.train --config ../configs/dermdepth_exp_g.json --workspace ../output/training/exp_g
# Stage 3: + DDI pseudo-GT (best model)
python -m moge.train --config ../configs/dermdepth_exp_h.json --workspace ../output/training/exp_h# Evaluate on SKINL2
python code/evaluation/eval_depth.py --model MoGe/pretrained_moge2.pt --dataset skinl2 --split test
# Evaluate on WoundsDB
python code/evaluation/eval_depth.py --model MoGe/pretrained_moge2.pt --dataset woundsdb --split test
# Evaluate on DDI (with ruler GT and fairness analysis)
python code/evaluation/eval_ddi_rulers.py --model MoGe/pretrained_moge2.pt --split test
# Evaluate baselines (DA3, MapAnything, PPD)
python code/evaluation/eval_baselines.py --method da3 --dataset skinl2This work builds on:
- MoGe-2 (Wang et al., 2025) for the base architecture
- S-SYNTH (Kim et al., MICCAI 2024) for the synthetic skin rendering framework
@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},
publisher = {Springer},
series = {Lecture Notes in Computer Science}
}See LICENSE for details.