Medical Foundation Models and Statistics for Detecting Distribution Drifts in Medical Data

Medical Foundation Models and Statistics for Detecting Distribution Drifts in Medical Data

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

OOD data is detected, among others, based on likelihoods, however their reliability is not always optimal.  In 2019, Likelihood Ratios for Out-of-Distribution Detection, it is shown that LRs can improve OOD detection reliability, assuming background statistics can be modeled accurately Medical image datasets are often unbalanced, with the background statistics prevailing in the majority of the data. For example, in the case of tumor detection, most samples are healthy, containing background statistics, whereas in anomalous samples (with tumors), the background statistics dominate. Normalizing flows and VAEs have been used to extract likelihoods for OOD detection, however Foundation Models may also be used for obtaining more accurate representations [1].

• You should adapt the method of [2[ on image segmentation using Medical Foundation Models.
• Compare OOD detection using likelihood and Likelihood Ratio.
• Assess on Medical Image benchmarks.

REFERENCES:

1. Your Finetuned Large Language Model is Already a Powerful Out-of-distribution Detector
2. Likelihood Ratios for Out-of-Distribution Detection
3. Revisiting Likelihood-Based Out-of-Distribution Detection by Modeling Representations