Recent research in Nature Biomedical Engineering highlights the limitations of artificial intelligence (AI) tools in detecting molecular biomarkers from histological images. These AI models often rely on correlational relationships between clinicopathological features, resulting in shortcuts that compromise their reliability in patient care. Study author Fayyaz ul Amir Afsar Minhas, PhD, analogizes this to judging a restaurant by its queue rather than food quality. The research analyzed over 8,000 tissue samples from various cancers, revealing that interdependencies among biomarkers can mislead predictions, as seen when AI conflated BRAF mutations with microsatellite instability. This underscores the need for rigorous, bias-aware AI evaluation to ensure accurate predictions and avoid erroneous treatment decisions. Researchers advocate for a stratification-based evaluation framework to enhance the development of trustworthy models in cancer diagnostics without replacing traditional molecular testing. Overall, caution is advised in utilizing AI tools, emphasizing their limitations while recognizing their potential in cancer research and treatment decision-making.
New Research Indicates AI Pathology Models Could Use Unreliable Shortcuts in Detecting Cancer Biomarkers
Share
Read more