SAC-Diff: A Scan-Aware Consistency-Enhanced Diffusion Framework for Unsupervised Chest CT Anomaly Detection

Xinyuan Zheng, Yoshihisa Shinagawa, Sepehr Farhand, Chi Liu, Gerardo Hermosillo, Xueqi Guo

Keywords: computer-aided diagnosis; unsupervised anomaly detection; CT

TL;DR: A scan-aware, consistency-enhanced diffusion model that improves unsupervised chest CT anomaly detection on heterogenous ILD and COVID datasets, achieving significant performance and robustness gains

Abstract: Anomaly detection in medical imaging is important but challenging due to diverse and imbalanced pathologies. Supervised methods rely on large annotated datasets and generalize poorly to unseen conditions. Unsupervised generative methods, especially diffusion models, can learn normal anatomy and detect outliers, but often hallucinate because of the Gaussian noise design and insufficient anatomical guidance. To address these challenges, we propose **SAC-Diff**, a **S**can-**A**ware **C**onsistency-enhanced **Diff**usion framework for unsupervised anomaly detection in automated lung disease screening using chest CT. SAC-Diff adopts simplex noise for detail-preserving diffusion perturbation, integrates scan awareness via (A) subject-aware anatomical priors into conditional diffusion and (B) background-aware masking for scan-specific variations and heterogeneous lung anomalies, and enhances robustness by enforcing consistency and quantifying uncertainty through multi-sample ensembling. We evaluate SAC-Diff on two diseased datasets with various anomalies, COVID-19 and interstitial lung disease (ILD), and observe substantial improvements over prior methods. On COVID-19, SAC-Diff achieves an IoU of 0.39 (+3.75% improvement compared to existing methods) and Dice of 0.53 (+2.99%); on ILD, it improves IoU to 0.31 (+74.45%) and Dice to 0.44 (+60.40%). Our results demonstrate promise toward robust and annotation-free CT anomaly detection in hospital deployment.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Visa & Travel: Yes

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: 351