Master Assignment
Detecting medical distribution shifts in hyperbolic spaces
Type: Master CS
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Distribution shifts between training and testing data can cause AI models to fail, often silently, producing overconfident, yet erroneous estimates. This causes issues in AI reliability, particularly in safety-critical applications, such as the analysis of medical data. Multiple methods have been developed for detecting out of distribution data, which are successful to a degree. For the most diAicult case of distribution shifts, namely near-OOD data, methods that are based on the Euclidean distance cannot easily discern between IN and OOD data.
Recent advances in hyperbolic learning1, 2 show a promising direction that may address some of the issues faced in OOD detection, as distances that are very small in Euclidean spaces, such as in the case of near OOD, may become large in hyperbolic space. Moreover, hyperbolic space may provide a better, and natural, representation of concepts that have a hierarchical structure.
In this project, you should investigate the performance of OOD detectors (such as Mahalanobis distance, softmax, energy scores) on medical image data when mapped from Euclidean to hyperbolic space. Your goals are to find:
- How well does OOD detection in hyperbolic space work for medical data?
- Focus #1 on medical datasets from the SoA.
- Focus #2 on real-world medical datasets (real-world ZGT hospital data is available for experiments on the hospital premises, in a privacy-preserving manner).
- What are the limitations in the most challenging case, when data is near-OOD?
- Focus on medical datasets that are near-OOD, i.e. very small distribution shifts. Examine wide range of datasets, i.e. from histopathology to CT scans, MRI for robustness and/or identification of limitations and advantages.
REFERENCES: