Master Assignment
[B] Benchmarking Medical Class-OOD (aka Open Set Recognition) Detection
Type: Master EE/CS/ITC
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Background:
The detection of Class-Out-Of-Distribution involves the detection of samples with classes that do not exist in the training data. For example, a model trained on lung Xray images receives a test image that is a knee or lung MRI.
Multiple methods exist for OOD detection. A popular one is based on Confidence Rate k, the Softmax score. It is based on the assumption that OOD samples will obtain a lower k score, however this is not always the case.
This thesis aims to study the behavior of confidence values, but also other OOD metrics, such as the Mahalanobis distance, for different cases of OOD data in benchmark medical image datasets.
Resources:
- A Framework For Benchmarking Class-Out-Of-Distribution Detection And Its Application To Imagenet (https://openreview.net/forum?id=Iuubb9W6Jtk)
- MOOD 2020: A Public Benchmark for Out-of-Distribution Detection and Localization on Medical Images (https://pubmed.ncbi.nlm.nih.gov/35468060/)