Research
Simplicity bias limits spectroscopic imaging-trained deep learning models for histology
Patterns · By Ji-Hun Oh, Dou Hoon Kwark, Kianoush Falahkheirkhah, Kevin Yeh, John Cheville, Volodymyr Kindratenko & Rohit Bhargava ·
The Brief
- Oh et al. report that deep learning tools built for RGB images can perform poorly when applied to chemically rich infrared images of tissue.
- On these infrared images, tumor-detection models learn only chemical signatures and largely ignore spatial information.
- The study is published in Cell Patterns.
Why it matters
Models designed for ordinary photos may take shortcuts on scientific imaging data, so researchers need to check whether their models use spatial structure rather than only chemical signals.
Who’s affected
- researchers in computational pathology and histology
- scientists applying deep learning to spectroscopic imaging
- #deep learning
- #histology
- #spectroscopic imaging
- #infrared imaging
- #simplicity bias
- #tumor detection
Deep learning tools built for RGB images are increasingly being applied to scientific data. Oh et al. show this can backfire. On chemically rich infrared images of tissue, tumor-detection models take a shortcut: they learn only chemical signatures, while largely ignoring spatial…