AI cancer pathology tools prone to shortcut learning instead of detecting true biology
University of Warwick researchers studied over 8,000 patient cancer samples and found that popular deep learning systems used in pathology may rely on statistical shortcuts rather than detecting true biological signals. When biomarkers do not occur together, these AI tools reportedly produce unreliable predictions, raising concerns about their safety for patient care.
What happened
Researchers from the University of Warwick analysed more than 8,000 patient samples across four major cancer types—breast, colorectal, lung and endometrial—and compared the performance of leading machine learning approaches used in cancer pathology.
The team found that while the models often achieved high headline accuracy, this frequently came from statistical shortcuts rather than genuine biological signals. For example, instead of detecting mutations in the BRAF gene, a model might learn that BRAF mutations often occur alongside microsatellite instability (MSI) and use this combination of cues to predict BRAF status rather than learning the causal BRAF signal itself.
Reported impact
When biomarkers do not co-occur, accurate cancer predictions become unreliable. The research raises concerns that some AI pathology tools are currently too unreliable for real-world patient care, as these shortcuts often fall apart when conditions change.
Index analysis
When performance of AI models was assessed within stratified patient subgroups, the team found these shortcut-based predictions failed. The research suggests many deep learning systems trained for cancer pathology may be relying on hidden shortcuts rather than genuine biological signals.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.