Keywords: synthetic data evaluation, network-induced artifacts, mammography
TL;DR: We propose a knowledge-based method to detect network-induced shape artifacts in synthetic images and demonstrate its use on synthetic mammography images.
Abstract: The adoption of synthetic medical images for training or testing without thorough quality assessment risks introducing artifacts and unrealistic features that can mislead machine learning models and compromise clinical utility. This work introduces a novel knowledge-based method for detecting network-induced shape artifacts in synthetic images. The method can identify anatomically unrealistic images, detect shape artifacts irrespective of the generative model, and offer interpretability through its knowledge-driven design. We validated the method using two synthetic mammography datasets and demonstrated its effectiveness in flagging images with network-induced artifacts. A reader study further confirmed these findings and showed that the most anomalous images identified by the method were also flagged by human readers. This method provides a step toward the responsible use of synthetic data by ensuring synthetic images align with realistic morphological and anatomical constraints.
Primary Subject Area: Generative Models
Secondary Subject Area: Safe and Trustworthy Learning-assisted Solutions for Medical Imaging
Paper Type: Methodological Development
Registration Requirement: Yes
Reproducibility: The code will be made available with the published paper. Three out of four datasets used in this work are public datasets. The fourth dataset was generated using the public implementation of the StyleGAN2 model without any modifications.
Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl25_NNN.zip file with midl25_NNN.tex, midl25_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2025}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \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)., Appendices and supplementary material are included in the same PDF after references., Main paper does not exceed 9 pages; acknowledgements, references, and appendix start on page 10 or later.
Submission Number: 124
Go to MIDL 2025 Conference homepage