Eike Petersen

Benchmarking saliency map faithfulness is so sensitive to perturbation design and class-conditioning that minor evaluation choices can completely flip which explainability method appears most trustworthy in medical ViTs.

Achieving top-tier medical image analysis performance with less than 2.5% of the parameters required for full fine-tuning could revolutionize accessibility in healthcare AI.

Medium-far and far-OOD samples can be detected reliably, but near-OOD anomalies expose critical method-specific vulnerabilities that challenge current practices in medical AI.

Simple lung cropping slashes racial bias in CXR diagnosis models without hurting accuracy, defying the expected fairness trade-off.