Publication: ESMO 2026

Authors: L. Van’t Veer; Laboratory Medicine, UCSF – University of California San Francisco – Parnassus Campus, San Francisco, United States of America

Title: AI Driven Digital Pathology for the 70 Gene Recurrence Signature: Development and Validation in Early Stage Breast Cancer

Abstract

Background: MammaPrint (MP) is a validated 70-gene recurrence risk assay used to inform treatment decisions in early-stage breast cancer (EBC). We explored whether the prognostic information captured by MP could be inferred directly from routine hematoxylin and eosin (H&E)–stained whole-slide images (WSIs) using artificial intelligence (AI).

Methods: An AI-based digital pathology model was trained on 70,000 H&E-stained WSIs with MP-derived supervision. The deep learning pipeline included tissue detection and tissue-type classification. Patch-level features were encoded using self-supervised and supervised foundation models. The approach generated a continuous recurrence risk score dichotomized into high- and low-risk categories. Retrospective clinical validation is performed on 376 patients from the RASTER study, which enrolled early-stage breast cancer patients, managed in a real-world shared decision-making setting (PMID 36126477). Assay concordance, and clinical performance are evaluated using distant recurrence free interval (DRFI) with 5- and 10-year follow-up data. Analyses include Kaplan–Meier with log-rank tests and multivariable Cox models adjusted for clinicopathologic factors.

Expected Results: The model is expected to accurately predict MP recurrence risk and demonstrate independent prognostic value across both 5- and 10-year follow-up intervals, including significant association with DRFI, clear risk group separation, and concordance with MammaPrint stratification.

Expected Conclusions: This study is expected to demonstrate AI-based 70-gene digital pathology as an easily accessible approach for recurrence risk stratification applied to routine H&E-stained WSIs for EBC.