Predicting Microvascular Decoupling and Osimertinib PFS from Routine H&E Whole-Slide Images via Attention-MIL (< 5 € / slide)
Democratizing spatial precision oncology: translating 5,000 € spatial transcriptomics into a routine diagnostic artificial intelligence framework operating directly on universal diagnostic H&E slides at under 5 € per patient.
Gated attention multiple instance learning architecture extracting complex microarchitectural sanctuary features from gigapixel whole-slide images.
Reduces the diagnostic cost barrier by 99.91% compared to spatial transcriptomics, enabling global deployment in public healthcare systems.
Achieves robust survival concordance (Harrell's C-index ≈ 0.74) for predicting osimertinib therapeutic durability directly from routine histology.
Lightweight streaming inference engine processing gigapixel WSI bags in Google Colab T4 environments without heavy storage overhead.