Normative Learning

Establishing the generative foundation for anomaly detection by modeling the healthy biological manifold.

Analyzing_Manifold

Scientific Thesis

"Normative learning provides the formal foundation for anomaly detection by establishing the high-dimensional 'healthy' state as a baseline. Only by accurately modeling this baseline can we identify pathology not as a pre-defined category, but as a statistically significant departure from a patient's expected anatomical manifold—enabling the discovery of novel disease markers through high-fidelity pseudo-healthy synthesis."

Open Challenges

Context-Aware Restoration

Architecting generative frameworks that resolve the fundamental tension between pathological correction and the preservation of patient-specific healthy anatomical context.

High-Dimensional Covariate Modeling

Conditioning generative priors on continuous variables—aging, sex, and acquisition hardware—to ensure the normative manifold is robust against non-pathological variance.

Multiscale Systems Integration

Augmenting structural imaging with molecular signatures and population-level insights to bridge the gap between microscopic biological drivers and macroscopic clinical outcomes.

Key Publications

Evaluating Normative Representation Learning in Generative AI
Nature Communications 2025 Full Manuscript

Evaluating Normative Representation Learning in Generative AI

Cosmin I. Bercea, Benedikt Wiestler, Daniel Rückert, Julia A. Schnabel
Diffusion Models with Implicit Guidance for Medical Anomaly Detection
MICCAI 2024 Full Manuscript

Diffusion Models with Implicit Guidance for Medical Anomaly Detection

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