Master Assignment
NLMedical OOD Segmentation Detection
Type: Master CS
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Personal pageOOD detection for cardiac MRI segmentation is examined in [1], on an IN-distribution dataset, the publicly available Automatic Cardiac Diagnosis Challenge (ACDC). The OOD images include artificially transformed in-distribution (ACDC) test images, cine cardiac MRI images acquired using different MR scanners or imaging protocols, cardiac images from different MR modalities, and non-cardiac images such as abdominal and lung CT scans and ADE20K scene-centric natural images. In this work a 2D U-Net is used to segment the CMR images.
As the paper says “In segmentation, the class distribution is often imbalanced, with the background class being much more prevalent than the objects of interest “, so they propose a new “Dice coefficient-based image-level uncertainty metric called Dice within samples (Dice-ws) to detect OOD samples. It is the
average Dice coefficient of the mean predicted segmentation 𝑆mean and the individual M prediction samples”. This new metric is shown to be less sensitive to dataset imbalance. Results from the paper give an average Dice score of 0.808 on the ACDC test dataset images, which are the in-distribution images. However, its performance decreased significantly on the near OOD datasets.
In this project you should implement and improve the post hoc OOD method for cardiac MRI segmentation.
- You should improve the segmentation in the pipeline by using other models such as nnUNet or Foundation Models like MedSAM, MedSAM2, MedSAM3, VISION-MAE.
- The FMs can be fine-tuned to further improve results
- Then test out the model e.g. compare the proposed Dice-ws metric and feature distance metrics on the
new, richer embeddings. - And/or you can use OOD detection libraries:
REFERENCES:
- Post-hoc out-of-distribution detection for cardiac MRI segmentation (2025)
- Research on Medical Image Segmentation Based on SAM and Its Future Prospects (2025)
- PromptSeg: Learning to Segment Medical Image via Visual Prompts (2025)
- PSAM: Prompt-based Segment Anything Model Adaption for Medical Image Segmentation (2024)
- Segment anything in medical images (2024)
- On the OOD Robustness of Foundation Models in Medical Image Segmentation (2023)
- Segment Every Out-of-Distribution Object (2024)
- PyTorch Out-of-Distribution Detection