Artificial Intelligence-Powered Prediction of Mohs Surgery Post-operative Appearance

Background

Mohs micrographic surgery involves complex intraoperative decision-making, making prediction of post-operative appearance challenging. Scarring has significant psychological impact on patients.^[1]

We present a personalized AI tool that gives patients a realistic preview of post-operative appearance and helps clinicians visualize different reconstructive approaches before surgery.

Stanford Mohs Surgery Dataset

We curated the largest longitudinal, matched clinical image dataset of Mohs surgery to date, spanning 2012–2025 across 8,033 patients.

Extensive annotation includes:

• Precise outlining of defects, wound closures, and scars
• Matching of intra-op, post-op, and follow-up images by patient, surgical site, and facial view
• Disambiguation of patients with multiple surgical sites across years of care
• Extraction of clinical metadata from surgery notes

Methods

We leverage Flux.1 Fill^[3], a state-of-the-art foundation model for image inpainting, to synthesize the surgical closure region while preserving the patient's unique facial features.

Key challenges addressed:

• View misalignment: Longitudinal image pairs help the model recognize different views of the same patient across visits.
• Wound closure localization: Inflated and randomly offset square bounding boxes reduce model bias on orientation and location of wound closure.

Results

Our model produces realistic and controllable post-operative predictions across a wide range of reconstruction types.

Quantitative Analysis

• n = 30 image pairs (AI-generated vs. real post-operative outcomes)
• 3 board-certified dermatologists in a blinded discrimination task
• Accuracy: 48.9% (p = 0.916)
• Fleiss' κ = 0.022

Supported reconstruction types:

• Linear repair
• Rhombic transposition flap
• Full-thickness skin graft
• Bilobe transposition flap
• V-to-Y advancement flap
• Trilobe transposition flap
• O-to-T advancement flap
• Healing by second intention

Additional reconstruction types in progress:

• Paramedian forehead flap
• Melolabial interpolation flap
• Perialar crescentic advancement flap
• Mucosal advancement flap
• Bilateral O-to-Z rotation flap
• Retroauricular interpolation flap
• Lip wedge repair
• Rieger dorsal nasal flap
• Purse string repair

Qualitative Results

We compare our model against state-of-the-art generative models on a standardized prompt: "Generate a photorealistic image of a Mohs micrographic surgery linear closure."

Facial images are not shown here to protect patient privacy.

Future Work

• Model enhancement: Improve fine-grained controllability of generated reconstructive approaches.
• Patient integration: Conduct clinical trials to evaluate patient perception and psychological impact when previewing individualized predictions prior to informed consent.
• Expansion: Apply this generative pipeline to other types of skin image data and procedures.

Funding

This work was supported in part by the Stanford Institute for Human-Centered Artificial Intelligence (HAI).

References

1. P. Kamath, C. Kursewicz, G. Ingrasci, R. Jacobs, N. Agarwal, and K. Nouri, "Analysis of patient perceptions of Mohs surgery on social media platforms," Arch Dermatol Res, vol. 311, no. 9, pp. 731–734, Nov. 2019, doi: 10.1007/s00403-019-01944-7.
2. W. C. Fix, C. J. Miller, J. R. Etzkorn, T. M. Shin, N. Howe, and J. F. Sobanko, "Comparison of Accuracy of Patient and Physician Scar Length Estimates Before Mohs Micrographic Surgery for Facial Skin Cancers," JAMA Network Open, vol. 3, no. 3, p. e200725, Mar. 2020, doi: 10.1001/jamanetworkopen.2020.0725.
3. O. Greenberg, "Demystifying Flux Architecture," arXiv, Jul. 2025, arXiv:2507.09595v1 [cs.CV].