SADUSI: multi-source ultrasound benchmark shows self-supervised anomaly detection still struggles

Diffusion-based detectors reach only 0.56 to 0.72 pixel-level AUROC and PatchCore variants 0.76 to 0.83 when anatomy, views, and protocols vary.

Primary source: arXiv:2610.09677: https://arxiv.org/abs/2610.09677

What’s new: Self-supervised anomaly detection is attractive for ultrasound because normal images are easier to collect than exhaustive pathology annotations. Most evaluations, however, cover a single anatomy or task. A Multi-Source Ultrasound Benchmark Revealing the Limits of Contemporary Self-Supervised Anomaly Detection Methods (arXiv:2610.09677, 7 Oct 2026) introduces SADUSI, a multi-source dataset spanning many anatomical regions, views, and acquisition protocols, with a diverse normal distribution for training and a benchmark of visible structural anomalies assessed from single images.

On this benchmark, reconstruction-based diffusion methods (AnoDDPM, DeCo-Diff) reach pixel-level AUROC of 0.56 to 0.72 and maximum F1 of 0.10 to 0.26. Feature-based PatchCore variants do better, with pixel-level AUROC of 0.76 to 0.83, but maximum F1 stays between 0.14 and 0.40. The authors conclude that broad multi-source ultrasound anomaly detection is an open problem.

Why it matters: Single-anatomy results can overstate how well a method has learned a general notion of normal appearance. A benchmark that varies anatomy and protocol exposes whether models learn source-specific representations, and the large gap between AUROC and F1 is a useful reminder that threshold-free metrics can look acceptable while localization remains poor.

Caveats: The benchmark targets visible structural anomalies that can be judged from a single image, so it does not cover temporal or functional findings. Only representative existing methods are evaluated, and the paper does not propose a new detector. Results reflect these specific methods and training setups, not a ceiling for the approach.