The Unseen & Unknown

Identifying rare clinical phenotypes and resections without expert labels through OOD detection.

Anomaly_Detected

Scientific Thesis

"Unsupervised Anomaly Detection (UAD) is formulated as a label-free approach for open-world clinical settings, where the space of possible pathologies is inherently unbounded, yet is predominantly developed and evaluated under closed-world assumptions tied to known anomaly distributions. We reframe UAD as a problem of generalizing anomaly discovery by grounding detection in principled modeling of the normative anatomical manifold, enabling reliable identification of rare, heterogeneous, and previously unseen disease phenotypes."

Open Challenges

Agnostic Sensitivity

Develop representations that remain sensitive to the full pathological spectrum without encoding assumptions about anomaly appearance, prevalence, or intensity.

The False Positive Bottleneck

Disentangle pathological deviations from benign anatomical variability to prevent increased open-world sensitivity from translating into clinically prohibitive false-positive rates.

Integrity of Open-World Generalization

Establish datasets and evaluation protocols that explicitly stress test UAD methods under open-world conditions, including rarity, heterogeneity, and previously unseen disease phenotypes.

Key Publications

NeurIPS 2025 (Oral) Full Manuscript

NOVA: A Benchmark for Rare Anomaly Localization and Clinical Reasoning in Brain MRI

Cosmin I. Bercea, Jun Li, Philipp Raffler, Evamaria O. Riedel, Lena Schmitzer, Angela Kurz, Felix Bitzer, Paula Roßmüller, Julian Canisius, Mirjam L. Beyrle, Che Liu, Wenjia Bai, Bernhard Kainz, Julia A. Schnabel, Benedikt Wiestler

→

MIDL 2023 Full Manuscript

Generalizing Unsupervised Anomaly Detection: Towards Unbiased Pathology Screening

Cosmin I. Bercea, Benedikt Wiestler, Daniel Rückert, Julia A. Schnabel

→