This workshop will explore how artificial intelligence is used to reconstruct biomedical microscopy data, emphasizing both the opportunities and the risks of these approaches. While AI has enabled remarkable advances in virtual staining, super-resolution, and image translation, biomedical imaging presents unique challenges that make hallucinations far more consequential than in other domains. A false structure in a pathology slide is not just an artifact, it can represent a disease feature or even a “patient” that never existed. The workshop will therefore focus on the nuances of model selection, data preprocessing, and evaluation strategies needed to responsibly apply AI in biomedical microscopy.
Target Audience: This workshop is designed for computational scientists collaborating with biomedical researchers who want deeper insight into the challenges of high‑stakes medical imaging. It is also open to undergraduate and graduate students, as well as other biomedical researchers.
Learning outcomes:
- Identify risks of AI hallucinations in biomedical imaging and explain why these differ from other application areas.
- Compare the performance characteristics (strengths and weaknesses) of GANs, CycleGANs, and diffusion models in microscopy reconstruction.
- Evaluate the role of data cleaning, preprocessing, and representation in reducing model failure.
- Apply criteria to assess whether an AI-generated biomedical image is trustworthy for downstream interpretation.
- Differentiate between appropriate and inappropriate use cases of AI in microscopy based on model behavior and validation evidence.
This workshop will be hosted by IDRE Fellow, Dr. Paloma Casteleiro Costa.
