Hyperbolic VAEs for better detection of out of distribution medical images

Hyperbolic VAEs for better detection of out of distribution medical images

Type: Master CS

Period: TBD

Student: (Unassigned)

If you are interested please contact :

dr. A. Briasouli PhD (Alexia) Assistant Professor

+31534896985

 a.briassouli@utwente.nl

Building: Zilverling 4049

Personal page

Latent spaces are compressed internal representations of data, often learned by models like Autoencoders: they compress raw data, and often capture useful interpretable information. Latent variables may naturally correlate with clinical features such as gender, organ volume, age – even when training is unsupervised1. Latent spaces may also encode contrast diAerences, scanner artifacts, noise. Thus, if latent spaces cluster well, they can be useful for detecting outliers, unusual anatomies, rare diseases, scan corruptions. But! In practice their separability is not always guaranteed1, 2. 

Latent spaces better cluster under supervised models, while unsupervised models like VAEs may organize data along smooth variations (e.g. variations in morphology, intensity etc). Hyperbolic geometry may help in this direction, as it naturally disentangles, while its geometry increases the distances of near OOD data from IN data, compared to Euclidean space3, 6. In MICCAI 2025, the first paper on anomaly detection in hyperbolic latent spaces was published, with
good results7 on medical benchmarking datasets. These results are promising and show potential.

• In this project, you should extend and deepen this work, with a more thorough investigation than in 7 of the eAects of using hyperbolic spaces on medical data. There are several ways to go about this, using diAerent implementations of hyperbolic VAEs (3, ,4, 5, 6, 7 below). You should test them out on downstream tasks of your choice. 
• The downstream tasks can range from simple classification to segmentation, anomaly detection and OOD detection.

REFERENCES:

1. Interpreting Latent Spaces of Generative Models for Medical Images using Unsupervised Methods (2022)
2. A study on the clusterability of latent representations in image pipelines (2023)
3. Hyperbolic VAE via Latent Gaussian Distributions
4. Learning Multi-Manifold Embedding for Out-Of-Distribution Detection
5. Hyperbolic latent VAE
   1. Github of a Hyperbolic VAE implementation applied to CelebA
   2. Can be used to reconstruct medical data + perform anomaly detection (as with a normal VAE)
6. Hyperbolic Metric Learning for Visual Outlier Detection
7. Is Hyperbolic Space All You Need for Medical Anomaly Detection?
8. Hyperbolic Anomaly Detection