Do Foundation Models See Biology? Evaluating Attention Coherence with Spatial Transcriptomics in Glioblastoma
Researchers have developed a new framework to evaluate whether foundation models used in pathology can accurately interpret biological data. This method uses spatial transcriptomics to assess the attention maps of five different pathology foundation models and a ResNet50 baseline. The findings indicate that these models capture complex transcriptional programs rather than individual molecular events, and that different models focus on distinct biological areas. AI
IMPACT Provides a quantitative method to assess AI model interpretability in pathology, crucial for clinical trust and regulatory approval.