Reproduce “Multi-modal AI for comprehensive breast cancer prognostication” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jan Witowski, Ken G. Zeng, Joseph Cappadona, Jailan Elayoubi, Khalil Choucair, Elena Diana Chiru, Nancy Chan, Young-Joon Kang and 43 more

The authors present a model combining visual features from H&E breast-tumor slides with clinical records to estimate recurrence risk. It was evaluated in five held-out cohorts spanning breast-cancer subtypes. For disease-free interval, its pooled C-index was 0.71 [0.68–0.75] across 3,502 patients. Among 858 Oncotype-tested patients, the estimate was numerically higher than Oncotype DX’s, but did not establish superiority.

Key insights

  • Self-supervised learning with a pan-cancer Vision Transformer (Kestrel) effectively extracts comprehensive prognostic features from digital pathology images.
  • Integrating AI-derived pathology features with standard clinical data produces a more accurate and robust prognostic tool than either modality or current genomic assays.
  • The multi-modal AI test provides consistent prognostic value across diverse breast cancer subtypes, including those without widely recommended molecular prognostic tools like TNBC and HER2+.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.