While deep networks have achieved broad success in analyzing natural images, when applied to medical scans, they often fail in unexpected situations.

Deep learning practice is increasingly driven by powerful foundation models (FM), pre-trained at scale and then fine-tuned for specific tasks of interest.