Canary CREST · Stanford Medicine · 2026

Physics-Guided Synthetic cNF Progression
Improves Image-Based Burden Estimation

Volume is all you need

Authors & affiliations 25 authors · 20 affiliations
Christoph Sadée1,*, Alex Dils1,2,*, Krish Sangani2, Lillian Rubino1, Alexander Ferenchick1,3, Varun Wadhwa4, Anushka Poddar5, Tobenna Onyemeh6, Zimuzo Onah7, Dhruv Sharma4, Jackson Bae8, Max Van Puyvelde1, Carlos Romo9, Melinda Jen10, Nkiru Onodugo11, Ayesha Akinkugbe12,13, Qinmei Xu1, Qingtao Kong14, Rui Yang15, Aisha Sokunbi16, Shaoxiong Yao17, Haomiao Huang18, Ifeoma Okoye6,11,19, Olivier Gevaert1,**, Kavita Sarin20,**
  1. Division of Computational Medicine, Department of Medicine, Stanford University
  2. University of California, Berkeley
  3. University of Pennsylvania
  4. University of California, Davis
  5. University of Maryland, College Park
  6. University of Nigeria, Nsukka, Center for Excellence for Clinical Trials
  7. University of Nigeria Teaching Hospital Ituku-Ozalla
  8. University of California, Los Angeles
  9. Johns Hopkins School of Medicine
  10. Children’s Hospital of Philadelphia, Dermatology Section
  11. University of Nigeria Teaching Hospital
  12. University of Lagos
  13. Lagos University Teaching Hospital
  14. Department of Dermatology, Jinling Hospital, Affiliated Hospital of the Medical School of Nanjing University, Nanjing, China
  15. Department of Dermatology, Second Affiliated Hospital of Nanjing Medical University, Nanjing, China
  16. University of Benin Teaching Hospital
  17. Siebel School of Computing and Data Science, University of Illinois Urbana-Champaign
  18. Matter Venture Partners
  19. University of Nigeria, Nsukka
  20. Department of Dermatology, Stanford University School of Medicine

* Co-first authors; ** Co-senior authors.

Fine-tuned Depth Pro reconstructs 3D skin surfaces from single images.

Abstract

Neurofibromatosis type 1 (NF1) is a tumor-predisposition syndrome, and the cutaneous neurofibromas (cNFs) that appear on the skin are its most visible sign. Following these lesions over time is clinically meaningful, yet the usual tools - visual inspection, manual counts, calipers, and flat photographs - describe how a lesion looks rather than how much tissue is actually there.

We connect that measurement problem to the missing time dimension. A morphoelastic finite-element model proposes how a lesion grows in three dimensions, and a geometry-conditioned FLUX.2 Klein adapter renders the simulated geometry as a realistic longitudinal image. Synthetic augmentation then improves burden estimation on held-out real images, increasing R² from 0.775 to 0.849.

Why volume?

cNFs are peripheral nerve-sheath tumors that grow within the skin itself, and a single person carrying NF1 can host hundreds or even thousands of them. The complication, for anyone hoping to measure them cleanly, is that they stubbornly refuse to be the same shape. They run along a whole morphological gradient: from nascent lesions barely lifted off normal skin, to flat ones that spread sideways without rising, to sessile ones planted broadly on the surface, to globular ones bulging outward, all the way to pedunculated ones dangling from a narrow stalk like fruit on a branch.

cNF morphology classes from nascent through pedunculated
cNF morphology classes, from nascent through pedunculated, with matched clinical photographs.

This variety is precisely what makes recent efforts in AI-driven cNF assessment so valuable and yet, on their own, incomplete. Detection and segmentation can answer where the lesions are and how many there are, but they stop short of the final step: total burden.

Clinically visible burden is a tangled product of count, footprint, protrusion, and morphology, and no one of those four ever stands in for the others. Two lesions can share an identical segmentation footprint while carrying very different amounts of tissue, one lying flat and the other bulging outward. Volume becomes the natural quantity to report because it folds together what the clinician can already see with what the 2D segmentations were hiding all along.

Volume is the measurement dermatology AI has been missing for cNF. It offers a lower-cost, faster way to quantify burden and a more precise way to follow disease over time. The need is also acute in research: cNF treatment trials still lack standardized outcome measures for objective changes in lesion size and appearance.

Method

A physics-to-image pipeline

A flat photograph gives us appearance, but progression requires geometry and time. The method therefore separates the problem in two. Physics supplies the shape a lesion should take as it grows; generative AI supplies the color, texture, and boundary detail that make the simulated state look like skin. The journey from one ordinary image to a longitudinal sequence runs through four steps.

Mechanistic modeling, AI rendering, reader-study evaluation, and volume-regression evaluation
Geometry is simulated first, then rendered as a realistic longitudinal image.
Step 1 · Encode

Recover lesion geometry

Depth Pro and surface normals describe how each lesion protrudes from the surrounding skin.

Step 2 · Simulate

Grow the lesion in 3D

A morphoelastic finite-element model expands one lesion through layered skin tissue.

Step 3 · Render

Turn geometry into skin images

A cNF-specific FLUX.2 Klein adapter renders each simulated geometry as a photorealistic frame.

Step 4 · Evaluate

Test on real held-out images

Reader studies assess realism, while matched ConvNeXt models test whether synthetic growth improves burden regression.

Synthetic Data

The longitudinal-data bottleneck

The pipeline sounds clean on paper: recover geometry, simulate growth, render the result. What makes the problem difficult is the missing ground truth. Public cNF datasets are almost entirely cross-sectional, and directly observing meaningful progression can require years or decades. A model cannot learn a time course from images that never contain one.

The way out is to manufacture the supervision ourselves. Depth and surface normals encode the protruding lesion. A layered finite-element skin model grows that lesion through a sequence of controlled states. FLUX.2 then converts each state back into a clinical-looking photograph while preserving the surrounding skin. What the generator hands us is exactly what ordinary photographs almost never can: paired images and geometry at multiple known growth states.

This is not synthetic data for its own sake. It is a controlled sandbox for building burden estimators, treatment-response tools, and medical-digital-twin infrastructure before scarce longitudinal patient data are available. Real patient imaging remains the standard for final calibration and validation.

An original cNF image followed by four simulated growth stages and a simulated scar
One observed image becomes four simulated growth states and a post-operative scar.

Data

Built from public images

279

cNF images

42,305

lesion annotations

279

generated sequences

The image split was 229 / 25 / 25 for training, validation, and testing. Scar modeling used 11 images with 359 annotations.

Results

What the synthetic data actually showed

The central question is whether the simulated growth states actually help on images the model has never seen. To isolate that effect, both training arms used the same architecture and the same 60 real images. The augmented arm received one additional ingredient: 240 physics-guided synthetic growth states.

Predicted versus target burden for real-only and real-plus-synthetic models
Three-seed ensemble evaluated on 40 held-out real images.

0.775 → 0.849

18.1%

lower log RMSE

The real-only arm used 60 training images. The augmented arm used the same 60 images plus 240 simulated growth states. On 40 held-out real images, RMSE fell from 1.179 to 0.965.

Fine-tuning also improved depth models

Depth-model RMSE before and after fine-tuning across four models
Average RMSE fell from 47.30 cm to 12.83 cm, a 72.9% reduction.

Limits

Promising, not clinically validated

  • Real 3D and longitudinal cNF images are still needed to validate the simulations.
  • The mechanistic model simplifies lesion biology and currently grows one lesion at a time.
  • The bootstrap intervals for the burden gain include no improvement, so the result remains uncertain.

Resources

Poster, paper, and code