Thesis
Learning Normative Anatomy under Limited Supervision
Doctoral Thesis (Dr.-Ing.), Friedrich-Alexander-Universität Erlangen-Nürnberg · 2026
Medical image analysis faces a structural asymmetry because the abnormalities a model must detect are open-ended, while the annotations it can learn from are scarce, costly, and disputed by experts who disagree on what they see. This thesis responds by learning normative anatomy, a model of what is normal, from limited supervision, and by treating that model’s uncertainty as a clinically actionable output rather than an afterthought. Viewed as a pipeline, anomaly detection embeds an assumption about what counts as normal at every stage; the thesis intervenes at three of them: how the training signal is synthesised, how deviation is scored, and how error is evaluated. The contributions span the range of label availability. With no anomaly labels, the training signal must be manufactured, i.e., self-supervised synthesis is recast as a stochastic process, yielding a distribution of plausible anomalies that mirrors annotator disagreement and improving localisation on chest radiographs and CT while halving training time (Probabilistic Poisson Image Interpolation). Given a normative segmentation model, anomalies are instead read off its uncertainty, fusing aleatoric and epistemic estimates over a frozen ultrasound foundation encoder so that unseen congenital heart defects surface as calibrated, localised uncertainty in real time (L-FUSION). That scoring can also be cheap and a backpropagation-free network can carry an intrinsic, interpretable anomaly score at a fraction of the usual compute and adapts to the resources at hand (SaFF-AD). At the defined-but-rare end of the spectrum, where the target is known but positive cases are few, an asymmetric similarity replaces the symmetric decision boundary (Similarity Fields). Underlying all of them is a generative model that supplies the normal prior they presuppose, producing clinician-validated healthy anatomy where that distribution was previously unavailable and releasing it as a privacy-safe dataset (SynthUterus). These contributions share as response to the opening asymmetry that abnormality is openended and labels are scarce, the dependable thing to model is normality, and its uncertainty is information the clinic can use. Pursued across the supervision spectrum, this reframes annotator disagreement as signal, replaces fixed decision boundaries with learned normality, and fits the compute budgets of real deployment. The result is a coherent set of self-supervised, uncertainty-aware, and efficient methods that bring medical anomaly detection closer to trustworthy clinical use.
@phdthesis{muller2026learning,
title = {Learning Normative Anatomy under Limited Supervision},
author = {M{\"u}ller, Johanna P.},
school = {Friedrich-Alexander-Universit{\"a}t Erlangen-N{\"u}rnberg},
year = {2026}
}