Flow Matching for 3D Craniofacial Skeletal Data Generation

Giacomo Melacini, Stefano Mazzocchetti, Giuseppe Lisanti, Luigi Di Stefano, Samuele Salti

Keywords: deep learning, generative modelling, generative, medical, 3D, flow matching, fm, diffusion models, ddpm, head CT, skulls, synthetic data, clinical task

TL;DR: 3D craniofacial skeletal data generation using Optimal Transport Flow Matching. The results are compared with DDPM and tested in two clinical downstream tasks.

Abstract:

In the medical domain, the use of Machine Learning (ML) techniques for diagnosis, treatment planning, and medical imaging interpretation is becoming increasingly important. However, these approaches require a large amount of data, which is challenging to access due to its sensitive nature and related privacy concerns. Synthetic data generation, enabled by advances in generative techniques, provides a solution to create large anonymized datasets for training models without compromising patient privacy.
Recently, Flow Matching with Optimal Transport (OTFM) has proven to be an effective technique for generating realistic 2D natural images, surpassing existing methods, but its usage for 3D medical data generation is limited.
In this work we generate craniofacial skeletal data using OTFM and test the validity of the results in two clinical downstream tasks: skull alignment and shape completion. Moreover, we compare the quality of synthetic data generated with OTFM with the ones generated using Denoising Diffusion Probabilistic Models (DDPMs). We show that Flow Matching with Optimal Transport is an effective technique for generating synthetic data and that, in this context, it outperforms DDPMs both in quality and robustness.

Primary Subject Area: Generative Models

Secondary Subject Area: Application: Neuroimaging

Registration Requirement: Yes

Reproducibility: https://github.com/Chavelanda/skeletal_fm

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 82