I have seen a scattering of articles that study whether AI interpretations of H&E slides can mirror the prognostic value of molecular tests.
The linked PDF white paper is entirely machine generated.
Here were the steps:
- I asked it to seek literature on the topic from PubMed.
- It got about 8.
- I asked it to download all the (public access) PDFs in a zip file for me.
- I then gave the PDFs back to it, and asked it to write a review article.
- I gave it about 8 or 10 topics or talking points to include.
- It produced a 12-page review.
- I asked it to produce an interesting report cover.
##
Find the white paper in the cloud here.
##
Here's the executive summary:
From Oncotype DX to “Virtual Recurrence Scores”:
AI Histopathology for Breast Cancer Risk Stratification
Abstract
Oncotype DX is the best-established multigene assay for guiding adjuvant chemotherapy in hormone receptor-positive, HER2-negative early breast cancer. Its evidence base spans retrospective validation, TAILORx, RxPONDER, and major guidelines, but cost, turnaround time, access, and tissue consumption remain limitations. A rapidly expanding literature now uses routine H&E whole-slide images, often combined with clinicopathologic data, to approximate Oncotype DX or related transcriptomic scores and, increasingly, to predict recurrence and chemotherapy benefit directly.
Recent studies show multicenter generalization, outcome discrimination, and analytical reproducibility. Yet score imitation, prognosis, treatment-benefit prediction, and clinical utility remain distinct claims and they will require separate validation before AI could replace genomic testing.
Recent studies show multicenter generalization, outcome discrimination, and analytical reproducibility. Yet score imitation, prognosis, treatment-benefit prediction, and clinical utility remain distinct claims and they will require separate validation before AI could replace genomic testing.
