# OpenDerm

> Source: https://aiwiki.ai/wiki/openderm
> Updated: 2026-07-31
> Categories: Healthcare AI, Open Source AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**OpenDerm** is an open-source [robotic](/wiki/robotics) imaging system that photographs human skin at close range with a four-degree-of-freedom gantry, then reconstructs and registers those photographs into a measurable three-dimensional map that can be compared across scanning sessions. It was designed and built by Marion Lepert, a robotics researcher who completed her PhD at [Stanford University](/wiki/stanford_university) in 2025, who published the hardware designs, software, and build documentation on 2026-07-27 and announced the project on X on 2026-07-29. [1][5][9] The stated purpose is longitudinal surveillance of moles and other skin lesions for early signs of melanoma, but the project describes itself explicitly as research equipment: "OpenDerm is research software, not a medical device, and does not diagnose melanoma or any other condition." [5]

| Field | Value |
| --- | --- |
| Type | Open-source robotic gantry for close-range skin imaging |
| Creator | Marion Lepert |
| Announced | 2026-07-29 (repository published 2026-07-27) |
| Degrees of freedom | 4 (linear X, Y, Z plus rotary RX pan) |
| Axis travel | X 0 to 800 mm, Y 0 to 665 mm, Z 0 to 392 mm |
| Working distance | 110 mm |
| Camera | Canon EOS R7 with Canon RF 100 mm macro lens |
| Illumination | Godox MF-R76 macro ring flash with cross-polarizing film |
| Distance sensing | Two Panasonic HG-C1100-P laser displacement sensors |
| Capture resolution | 32.5 megapixels per frame, about 13 um per pixel |
| Output texture resolution | 78 pixels per millimetre at full quality |
| Parts cost | $5,010.71 core build plus $3,122.75 camera and imaging, $8,133.46 total |
| Software licence | MIT |
| Hardware licence | CERN-OHL-P-2.0 (per the project website) |
| Regulatory status | None. Not a medical device; research use only |

## Background: the screening problem OpenDerm targets

Melanoma outcomes depend heavily on the stage at diagnosis. The US National Cancer Institute's SEER program reports five-year relative survival of 100.0% for localized melanoma of the skin, 76.0% for regional disease and 34.0% once it has metastasized, from SEER 21 data (excluding Illinois) for 2016 to 2022; 77% of cases are caught while localized, and the American Cancer Society projects 112,000 new US cases and 8,510 deaths in 2026. [10]

The advice to watch existing moles captures only part of the problem. A meta-analysis by Pampena and colleagues, pooling 38 observational studies, estimated that 29.1% of melanomas arise from a preexisting nevus while 70.9% appear de novo on skin that previously looked normal. [11] Detecting a new millimetre-scale lesion among hundreds of benign spots is therefore a registration task, and Lepert frames it that way: comparing a patient's own photographs over time is defeated by changes in body position, camera angle, distance, lighting and skin deformation. [7]

Clinical total-body photography (TBP) exists to supply that missing baseline. Canfield Scientific's VECTRA WB360 records the entire exposed skin surface in one capture using 46 stereo vision pods (commonly described as 92 cameras) with both cross-polarized and non-polarized lighting. [19] FotoFinder's ATBM master adds video dermoscopy of individual lesions via its medicam handpiece; wide-field systems generally use either an array of fixed cameras or a single camera repositioned automatically between views. [20][1] A 2025 scoping review in JMIR Dermatology found the VECTRA system promising but concluded that evidence does not yet support replacing clinician assessment, naming high cost and room footprint as barriers. [18]

The clinical evidence is mixed in a specific way. In a randomized trial at a research hospital in Brisbane, Australia, running from April 2018 to October 2021, Soyer and colleagues added VECTRA WB360 3D TBP plus sequential digital dermoscopic imaging, delivered by teledermatology every six months, to usual care for adults at high risk of melanoma. Among 314 participants who completed all procedures (158 intervention, 156 control), 1,527 lesions were excised (905 intervention, 622 control) and 67 melanomas were confirmed, about 4% of excisions: 24 in the intervention arm against 43 in control. Mean lesions excised per person rose from 3.99 to 5.73 (P = .02). The authors concluded that adding 3D TBP and sequential imaging in a teledermatology setting without [artificial intelligence](/wiki/artificial_intelligence) increased the number and rate of excisions and biopsies. [12] A companion cost-effectiveness analysis by Lindsay and colleagues reported mean costs of $1,708 per participant over 24 months against $763 for usual care, identical quality-adjusted life-years of 1.84 in both arms, and a 0% probability of cost-effectiveness at a threshold of AU$50,000 per quality-adjusted life-year. [13]

Lepert's reading is that resolution, not the concept of longitudinal imaging, is the binding constraint: wide-field arrays trade detail for speed, so each lesion occupies only a small part of the frame and every flagged spot still has to be found on the patient and inspected with a dermatoscope. [7]

## Creator

Marion Lepert completed a robotics PhD at Stanford University, defending in September 2025, advised by Jeannette Bohg in the Interactive Perception and Robot Learning Lab; her personal homepage still describes her as a final-year PhD student; she also represented the United States in windsurfing at the 2016 Olympic Games. [8] Her published research is in [robot learning](/wiki/robot_learning) from cross-embodied data, not medical imaging: Phantom (CoRL 2025), which trains manipulation policies from human video alone by inpainting the human arm and compositing a rendered robot into the frame [23]; Masquerade (ICRA 2026); Shadow (CoRL 2024); and co-authorship on the DROID [robot manipulation](/wiki/robot_manipulation) dataset (RSS 2024) and TidyBot (IROS 2023). [8] The blog post accompanying OpenDerm opens with her own account of a suspicious mole and a biopsy that proved benign. [7]

## The prototype

OpenDerm is a rectangular gantry built mostly from 40-series aluminium extrusion. Three linear axes position a sensor head (X along the side rails, Y across the top beam, Z for camera height) and a fourth rotary axis, RX, pans the head to align the camera with the local skin surface. Configured travel is X 0 to 800 mm, Y 0 to 665 mm and Z 0 to 392 mm, inside a published mechanical envelope of 1000 by 900 by 600 mm. [1][5] The subject lies underneath the frame during a scan.

The sensor head carries a Canon EOS R7 mirrorless body with a Canon RF 100 mm macro lens, a Godox MF-R76 macro ring flash fitted with linear polarizing film to suppress specular glare through cross-polarization, and two downward-facing Panasonic HG-C1100-P laser displacement sensors. The laser readings drive the RX tilt and the Z height until the camera sits at its 110 mm working distance with the optical axis aligned to the local surface normal, and only then does the camera fire. [1][3][5]

Control is split across three computers. One Raspberry Pi runs Klipper on a BIGTREETECH Octopus Pro board for the X axis and bridges a Raspberry Pi Pico driving Y and Z; a second handles the CAN-controlled RX gimbal motor (a CubeMars AK45-36 quasi-direct-drive [actuator](/wiki/actuator)), the camera through the Canon EDSDK, and an ADS1115 converter reading the laser sensors; a workstation performs reconstruction and comparison. The software is [Python](/wiki/python) 3.11 or newer, and non-loopback motion services refuse to start without a shared control token. [5]

The published bill of materials itemizes 94 core-build parts totalling $5,010.71 and four camera items totalling $3,122.75, for $8,133.46 before general-purpose tools. The EOS R7 body ($1,549) and the RF 100 mm macro lens ($1,349) dominate the camera side; the two laser sensors are the largest core-build line at $325.60 each. [3]

## How the imaging pipeline works

### Capture

The robot scans in a serpentine raster of discrete imaging stations. At each one it settles the pose, captures a 32.5-megapixel macro frame sampling the skin at roughly 13 micrometres per pixel, and writes a synchronized JSON file with the station index, encoder-derived camera pose and both laser distances. Adjacent frames overlap by about 75%, so most skin appears in several images from slightly different viewpoints. [2]

### Match and optimize

Reconstruction is a constrained [structure-from-motion](/wiki/pose_estimation) problem rather than flat mosaicking, for four documented reasons. The encoders resolve position at the millimetre scale while the images resolve 13 micrometres, so raw poses are far too coarse to place frames directly; in the published example scan, optimized poses differed from encoder estimates by a median of 3.5 mm and by as much as 14.3 mm. Breathing shifts the skin between passes. Projecting a curved surface onto a plane would distort lesion dimensions. And the macro lens sees only about 8 degrees, so image-only estimates cannot separate small translations from small rotations and tend to drift. [2]

The pipeline detects around 6,000 SIFT keypoints per frame on CLAHE-equalized imagery, uses the approximate robot poses to predict which frames overlap and how far features should move between them, and rejects candidates inconsistent with that prior. RANSAC accepts an image pair only when at least 25 correspondences fit a single geometric transformation. Surviving matches are chained into tracks, and the solver alternates between triangulating landmarks with camera poses fixed and refining each camera's six pose parameters with landmarks fixed, keeping the encoder poses as soft constraints. Before optimization, four views of the same mole landed up to 3.7 mm apart; afterwards, pores and individual hairs align. [2]

Two details matter for reproducibility. Focal length is treated as a fixed, externally calibrated quantity (39,237 pixels at full resolution for the reference camera) because focal length and depth are nearly degenerate at this field of view, and focus must be locked manually because autofocus changes effective magnification by roughly 1% per frame at macro distance. [6]

### Comparing scans months apart

The system does not try to reconstruct the body's instantaneous shape. The documentation states that the diagnostic signal is the two-dimensional skin texture and the three-dimensional surface is only a scaffold to unwrap onto, so it should be smooth, stable and reproducible rather than instantaneously accurate. The texture map is parameterized in the gantry's own coordinate frame, so the same patch of skin lands at the same map coordinate every session. [6]

The `openderm-compare` tool resamples two scans onto a common canvas at 20 pixels per millimetre, intersects their coverage masks, converts reflectance to a melanin-like log map, and aligns them in tiers: a SIFT plus RANSAC similarity fit accepted only if it is near identity, then a constellation fit treating the moles themselves as fiducials, then gantry-only alignment flagged as low confidence. Moles are detected by a multiscale Laplacian-of-Gaussian response and labelled new, disappeared or stable, with new and disappeared judgments hard-gated to skin covered in both scans, and every verdict carries a tier of reliable, provisional or low-confidence. Size changes carry an error bar combining registration residual, detector jitter and a null floor, and count as significant only when the z-score exceeds 2.5 and the change exceeds 0.30 mm. [6] The public demonstration compares two registered scans in which two marker dots were added between sessions as controlled changes; the system flags both as new. [4]

### In plain terms

A dermatoscope shows one lesion in fine detail. A whole-body camera rig shows every lesion, coarsely. OpenDerm tries for both by moving one high-magnification camera over the whole surface, and the price is that thousands of narrow close-ups must be stitched into a map accurate enough that a 0.3 mm change in a mole is a real signal rather than a stitching error. The robot's joint encoders supply a rough guess of where each photograph was taken, and the optimizer corrects that guess using the skin's own pores, hairs and freckles as landmarks.

## Licensing and availability

The project website states that OpenDerm software is released under the [MIT License](/wiki/mit_license) and that the hardware design files, schematics, bill of materials and build documentation are released under the CERN Open Hardware Licence Version 2 Permissive (CERN-OHL-P-2.0). [1] The [GitHub](/wiki/github) repository at `MarionLepert/openderm` carries an MIT `LICENSE` file at its root and ships the Python packages `openderm` (hardware control and capture) and `skinmap` (registration, artifact detection and comparison), Klipper configuration, Pico firmware, calibration and collision-envelope scripts, a robot URDF with collision meshes, and a hardware-free test suite. [5]

As of 2026-07-31 the repository had over 100 stars and 7 forks, no open issues or pull requests, and no commits after the publication date. [5] There is no accompanying paper or preprint.

## Where OpenDerm sits among skin-imaging approaches

The project groups large-area skin imaging into three families and places itself in the middle. [1]

| Approach | Examples | Capture time | Detail | Relative cost |
| --- | --- | --- | --- | --- |
| Wide-field total-body photography | Canfield VECTRA, FotoFinder, DermSpectra, Neko | Seconds | Wide-field | High |
| Close-range robotic scanning | SquareMind, iToBoS, OpenDerm | Minutes | High resolution | Medium |
| Guided smartphone imaging | SkinIO, MoleMap, Miiskin | Manual | Variable | Low |

Two of the named robotic peers are substantial funded efforts. iToBoS ("Intelligent Total Body Scanner for Early Detection of Melanoma") ran from April 2021 to March 2025 under EU Horizon 2020, coordinated by the Universitat de Girona, at a total cost of about 12.04 million euros. [22] SquareMind, a Paris company, raised $18 million on 2026-04-27 in a round led by Sonder Capital for Swan, a robot arm that captures full-body dermoscopic imagery in minutes. [21] OpenDerm's claim is therefore not novelty of concept but that a comparable capture regime can be assembled from catalogue parts and released openly. [1]

## Relevance to dermatology AI

Dermatology was an early showcase for [deep learning](/wiki/deep_learning) in medicine: Esteva and colleagues trained a [convolutional neural network](/wiki/convolutional_neural_network) on 129,450 clinical images spanning 2,032 diseases (757 training classes) and reported performance comparable to 21 board-certified dermatologists on binary malignancy tasks. [14] The International Skin Imaging Collaboration (ISIC) has since anchored the field's public benchmarks.

The gap OpenDerm is aimed at is visible in the ISIC data itself. SLICE-3D, the training set for the 2024 ISIC challenge hosted on [Kaggle](/wiki/kaggle), consists of 15 mm by 15 mm crops centred on individual lesions and extracted from 3D total-body photographs; the release includes 401,059 crops under a non-commercial licence and a 217,477-crop permissive subset. [16] Its descriptor paper states plainly that most public skin cancer datasets are dermoscopic and are limited by [selection bias](/wiki/selection_bias) and a lack of standardization, and that the SLICE-3D crops are "comparable in optical resolution to smartphone images." [15] The large longitudinal-capable corpora are low resolution and the high-resolution corpora are pre-filtered to lesions a clinician already found suspicious. Neither directly supports the question a screening model would need to answer, which is what changed on this person since last time.

Skin-tone coverage is a second, well-documented weakness of the [training data](/wiki/training_data). Daneshjou and colleagues assembled the Diverse Dermatology Images set of 656 pathologically confirmed clinical images across skin tones and found that state-of-the-art dermatology models lost 27% to 36% of their ROC-AUC relative to their original test results, with all models performing worse on dark skin tones and uncommon diseases. [17] OpenDerm has published no data on performance across skin tones.

Lepert's argument is that repeated, standardized, high-resolution scans of the same person would be a different training signal from either existing family, and that producing such [datasets](/wiki/dataset) is part of the point of building cheap scanners. [7] That remains a hypothesis: no OpenDerm dataset has been released, and no model has been trained on OpenDerm output.

## The home robotics argument

Lepert's public framing goes beyond the prototype. In the announcement post she wrote that the path to scale "is not a dedicated screening robot in every household, it is to make skin screening one of the many useful things a general-purpose home robot can do," and the essay puts it as "the same robot that cleans your house should also scan your skin." [9][7] Her reasoning is that open-sourcing lowers the barrier to reproduction but most people will never build a dedicated imaging machine, whereas a [general-purpose](/wiki/embodied_ai) household robot could scan often, under consistent conditions, without an appointment. [7]

This is a position argument, not a demonstrated result. OpenDerm's capture regime depends on a rigid gantry with sub-millimetre repeatability and encoder poses good to a few millimetres, which is exactly what the reconstruction uses as its geometric prior; no published work shows that a mobile [service robot](/wiki/service_robot) can supply an equivalent prior while a person stands freely in a room. The claim connects OpenDerm to broader arguments about [smart home](/wiki/smart_home) automation and to healthcare robotics platforms such as [NVIDIA Isaac for Healthcare](/wiki/nvidia_isaac_for_healthcare), but it is a forecast.

## Limitations

The project is candid about its status. The repository warns that operating the robot involves impact, pinch, electrical and laser hazards, and that limits and emergency stops should be validated with nobody in the workspace before imaging a person. [5] The website footer reads "Not a medical device. Research use only." [1] There is no regulatory clearance, no clinical validation, and no peer-reviewed or preprint evaluation as of 2026-07-31.

The practical limits are substantive. The subject lies under the gantry, which the author identifies as the first thing to change in a version 2 (scanning a standing patient would be safer), alongside replacing the ballscrew-driven axes with belts for speed. [7] Scanning is sequential and takes minutes rather than seconds, the acknowledged trade for resolution. [1] Cross-visit comparison needs generous overlap and enough stable lesions to estimate repositioning; the documentation says the tool should abstain otherwise, and recommends same-scan null tests before interpreting any change. [6] Camera intrinsics must be calibrated externally: the repository ships calibration scripts for the RX pivot and floor depth, but its documentation states that OpenDerm "does not provide a camera-calibration utility". [5] Anatomical coverage is unproven: the published results page shows one cropped body region rather than a whole-body scan, with no data on scalp, folds or acral sites. All quantitative claims come from the project's own documentation and have not been independently reproduced.

## Significance

OpenDerm is best understood as a published existence proof rather than a product. It shows that the close-range robotic scanning approach pursued by venture-funded companies and multi-million-euro consortia can be prototyped from catalogue parts for well under $10,000, and it releases the awkward parts (the pose-constrained reconstruction, the gantry-anchored unwrap, the tiered comparison with explicit abstention) rather than only the mechanical design. For [AI in healthcare](/wiki/ai_in_healthcare) more broadly, its sharpest claim is about data rather than hardware: that the missing ingredient in dermatology [computer vision](/wiki/computer_vision) is not another classifier but repeated, standardized, high-resolution imagery of the same skin over time. Whether an [open-source](/wiki/open_source) gantry, a home robot, or a commercial scanner produces that data is unresolved.

## See also

- [Robotics](/wiki/robotics)
- [AI in healthcare](/wiki/ai_in_healthcare)
- [Computer vision](/wiki/computer_vision)
- [Convolutional neural network](/wiki/convolutional_neural_network)
- [Open source AI](/wiki/open_source_ai)
- [Embodied AI](/wiki/embodied_ai)
- [Service robot](/wiki/service_robot)
- [Surgical robot](/wiki/surgical_robot)

## References

1. "OpenDerm: Open-source robotic 3D skin imaging." OpenDerm project website, 2026. https://openderm.github.io/
2. "How OpenDerm builds a 3D skin map." OpenDerm project website, 2026. https://openderm.github.io/how-it-works.html
3. "Bill of materials." OpenDerm project website, 2026. https://openderm.github.io/bom.html
4. "Skin scan results: lesions measured and matched across two registered scans." OpenDerm project website, 2026. https://openderm.github.io/results.html
5. Lepert, Marion. "openderm" (source repository and README). GitHub, published 2026-07-27. https://github.com/MarionLepert/openderm
6. Lepert, Marion. "Skin scan registration: goal and method" (docs/skin-registration.md). GitHub, 2026. https://raw.githubusercontent.com/MarionLepert/openderm/main/docs/skin-registration.md
7. Lepert, Marion. "Why Catching Skin Cancer Early Is a Home Robotics Problem." Personal blog, July 2026. https://marionlepert.github.io/blog/openderm-robotics-problem.html
8. "Marion Lepert." Personal homepage (biography and publication list), accessed 2026-07-31. https://marionlepert.github.io/
9. Lepert, Marion (@marionlepert). Post announcing OpenDerm. X, 2026-07-29. https://x.com/marionlepert/status/2082512842742489258
10. National Cancer Institute. "Cancer Stat Facts: Melanoma of the Skin." SEER Program, accessed 2026-07-31. https://seer.cancer.gov/statfacts/html/melan.html
11. Pampena R, Kyrgidis A, Lallas A, Moscarella E, Argenziano G, Longo C. "A meta-analysis of nevus-associated melanoma: Prevalence and practical implications." Journal of the American Academy of Dermatology, 2017;77(5):938-945.e4. https://pubmed.ncbi.nlm.nih.gov/28864306/
12. Soyer HP, Jayasinghe D, Rodriguez-Acevedo AJ, et al. "3D Total-Body Photography in Patients at High Risk for Melanoma: A Randomized Clinical Trial." JAMA Dermatology, 2025;161(5):472-481. https://jamanetwork.com/journals/jamadermatology/fullarticle/2831500
13. Lindsay D, Soyer HP, Janda M, et al. "Cost-Effectiveness Analysis of 3D Total-Body Photography for People at High Risk of Melanoma." JAMA Dermatology, 2025;161(5):482-489. https://pubmed.ncbi.nlm.nih.gov/40136266/
14. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. "Dermatologist-level classification of skin cancer with deep neural networks." Nature, 2017;542(7639):115-118. https://pubmed.ncbi.nlm.nih.gov/28117445/
15. Kurtansky NR, D'Alessandro BM, Gillis MC, et al. "The SLICE-3D dataset: 400,000 skin lesion image crops extracted from 3D TBP for skin cancer detection." Scientific Data, 2024;11(1):884. https://pubmed.ncbi.nlm.nih.gov/39143096/
16. International Skin Imaging Collaboration. "ISIC Challenge Datasets: 2024 (SLICE-3D)." Accessed 2026-07-31. https://challenge.isic-archive.com/data/2024/
17. Daneshjou R, Vodrahalli K, Novoa RA, et al. "Disparities in Dermatology AI Performance on a Diverse, Curated Clinical Image Set." arXiv:2203.08807 (published in Science Advances, 2022). https://arxiv.org/abs/2203.08807
18. Baete F, Jakers AL, Delanoye E, Vande Velde N, Voet G. "3D Total Body Photography as a Promising Innovation for Early Skin Cancer Detection: Scoping Review." JMIR Dermatology, 2025;8:e68510. https://pmc.ncbi.nlm.nih.gov/articles/PMC12710984/
19. Canfield Scientific. "VECTRA WB360 Imaging System." Product page, accessed 2026-07-31. https://www.canfieldsci.com/imaging-systems/vectra-wb360-imaging-system/
20. FotoFinder Systems. "ATBM master: AI-supported total body mapping for the early detection of skin cancer." Product page, accessed 2026-07-31. https://www.fotofinder-systems.com/technology/medical-imaging/atbm-master/
21. The Robot Report. "SquareMind raises $18M for robotic dermatology platform." 2026-04-27. https://www.therobotreport.com/squaremind-raises-18m-dermatology-robotics-platform/
22. CORDIS, European Commission. "Intelligent Total Body Scanner for Early Detection of Melanoma (iToBoS), Grant agreement ID 965221." Accessed 2026-07-31. https://cordis.europa.eu/project/id/965221
23. Lepert M, Fang J, Bohg J. "Phantom: Training Robots Without Robots Using Only Human Videos." Proceedings of the 9th Conference on Robot Learning (PMLR), 2025;305:4545-4565. https://proceedings.mlr.press/v305/lepert25a.html