Thank you to my co-authors @bernhardkainz.bsky.social and C. Müller

yes, exactly. Ensembles of methods usually achieve the best performance for challenging problems like this.

I mean the hidden biases can be learned, not only the known parameters. One can check with a parametrised model if a new input conforms with all the biases in a specific downstream model. This does not necessary need to be ethnicity but could be a combination of latent factors.

finding a method that allows us to estimate how suitable a model is for a given patient... This solves only one side of the problem but bias is not necessarily 'bad'. After all, 'patient specific' means 100% bias towards one specific patient. Triage and patient stratification will be key in future.

- Multiple Instance Learning for Axillary Lymph Node Metastasis (DEMI)
x.com/Hanna_Paulal...

- Lesion-Level Data Augmentation for Segmentation (DALI)
arxiv.org/abs/2308.09026

- Probabilistic Self-Supervised Image Anomaly Localisation (UNSURE)
x.com/Hanna_Paulal...

- Synthetic Data Enrichment for Robust Image Segmentation in Kidney Transplant Pathology (DART)
arxiv.org/abs/2304.09534

- Neonatal Cortical Surface Reconstruction
arxiv.org/abs/2307.11870

- Unsupervised anomaly localisation with synthetic tasks
arxiv.org/abs/2307.00899

- Respiratory Motion Correction and Super-Resolution for 3D Abdominal MRI
link.springer.com/chapter/10.1...

- Robust Segmentation via Topology Violation Detection
link.springer.com/chapter/10.1...

- Cascaded Diffusion Models to generate echocardiogram videos
x.com/reynaud_h/st...