Prediction of OncotypeDX recurrence score using H&E stained WSI images

Shachar Cohen, Gil Shamai, Edmond Sabo, Alexandra Cretu, Iris Barshack, Tal Goldman, Gil Bar‐Sela, Alexander T. Pearson, Dezheng Huo, Frederick M. Howard, Ron Kimmel, Chen MayermedRxiv 2025 · DOI

TL;DR

Researchers have developed a deep learning model that predicts OncotypeDX breast cancer recurrence scores directly from standard, low-cost H&E tissue slides. By leveraging a massive pathology foundation model (Prov-GigaPath), the team achieved high accuracy and robust generalizability across international cohorts, potentially offering a faster, cheaper alternative to expensive genomic testing.

What Problem Does This Paper Address?

In early-stage breast cancer, the OncotypeDX 21-gene assay is the gold standard for determining whether a patient needs chemotherapy. By analyzing the expression of specific genes, it produces a Recurrence Score (RS) that estimates the risk of the cancer returning.

However, OncotypeDX has significant hurdles: it is expensive (thousands of dollars), requires shipping samples to specialized labs, and often takes weeks to return results. In low- and middle-income countries, these barriers often mean patients must make treatment decisions without this critical information. This paper addresses the need for a "digital proxy"—an AI tool that can extract the same prognostic information from the Hematoxylin and Eosin (H&E) stained slides that are already produced for every cancer patient as part of routine care.

Methods

The authors moved away from training a model from scratch, which often fails to generalize across different hospitals. Instead, they utilized a Foundation Model approach:

  1. Prov-GigaPath: They used a model pre-trained on a staggering 171,189 whole-slide images (WSIs) using Self-Supervised Learning (SSL). This allows the model to learn a rich "vocabulary" of pathology features without needing human labels.
  2. Architecture: The system uses a fixed transformer-based tile encoder to turn small patches of the slide into feature vectors. These are then fed into a transformer-based slide encoder, which was fine-tuned specifically to predict the OncotypeDX RS.
  3. Calibration: To handle the "domain shift" (differences in staining and scanning equipment between hospitals), the researchers employed histogram matching to calibrate the model’s output scores across different datasets.

Datasets

The study is notable for its geographic and institutional diversity, utilizing five independent cohorts:

  • Internal (Development): Carmel Medical Center and Haemek Medical Center (Israel).
  • External (Validation): Sheba Medical Center (Israel), University of Chicago Medical Center (UCMC, USA), and the Australian Breast Cancer Tissue Bank (ABCTB, Australia).
  • Scale: A total of 5,546 slides from 4,227 patients were analyzed.

Key Insights and Ideas

  • Morphology as a Mirror of Genetics: The core "aha" moment is the confirmation that complex genetic signatures leave visible "fingerprints" in tissue architecture—such as nuclear atypia, mitotic activity, and immune infiltration—that AI can quantify more consistently than the human eye.
  • The Power of Scale: By using a foundation model trained on an order of magnitude more data than previous studies, the authors achieved much higher generalizability. While earlier models saw performance drops when tested on new hospitals, this model remained stable across international borders.
  • Prioritizing Clinical Safety: The researchers didn't just aim for high average accuracy; they optimized for Negative Predictive Value (NPV). In a clinical setting, it is vital that if the AI says a patient is "low risk" (and thus can skip chemo), it is almost certainly correct.

Key Results

  • Strong Performance: The model achieved an Area Under the Curve (AUC) of 0.817 and 0.825 on the two external cohorts with available OncotypeDX scores.
  • Clinical Utility: Using a high-sensitivity threshold, the model identified roughly 22-25% of patients as low-risk with a sensitivity of 0.95–0.97 and an NPV of 0.96–0.97.
  • Prognostic Value: In survival analysis, patients the model flagged as "high risk" had significantly worse outcomes. On the UCMC cohort, the high-risk group had a Hazard Ratio of 4.1 for recurrence compared to the low-risk group.
  • Superior to Traditional Metrics: When compared against standard clinical variables (like Ki67 expression or tumor grade), the AI-predicted score was the most significant predictor of the genomic assay result.

Limitations and Future Work

  • The "Black Box" Problem: Like most deep learning models, it is difficult to explain exactly which morphological features the transformer is prioritizing, which can be a barrier to clinical adoption.
  • Multimodal Integration: While the AI score was powerful on its own, the authors noted that adding clinical data (like PR expression) improved results. Future work could focus on a more integrated "AI + Clinical" score.
  • Prospective Validation: While the retrospective results on five cohorts are impressive, a prospective clinical trial would be the final step to prove the model can safely replace or supplement genomic testing in real-time practice.

Summary updated

Abstract

Abstract The OncotypeDX 21-gene assay is a widely adopted tool for estimating recurrence risk and informing chemotherapy decisions in early-stage, hormone receptor-positive, HER2-negative breast cancer. Although informative, its high cost and long turnaround time limit accessibility and delay treatment in low- and middle-income countries, creating a need for alternative solutions. This study presents a deep learning-based approach for predicting OncotypeDX recurrence scores directly from hematoxylin and eosin-stained whole slide images. Our approach leverages a deep learning foundation model pre-trained on 171,189 slides via self-supervised learning, which is fine-tuned for our task. The model was developed and validated using five independent cohorts, out of which three are external. On the two external cohorts that include OncotypeDX scores, the model achieved an AUC of 0.825 and 0.817, and identified 21.9% and 25.1% of the patients as low-risk with sensitivity of 0.97 and 0.95 and negative predictive value of 0.97 and 0.96, showing strong generalizability despite variations in staining protocols and imaging devices. Kaplan-Meier analysis demonstrated that patients classified as low-risk by the model had a significantly better prognosis than those classified as high-risk, with a hazard ratio of 4.1 (P<0.001) and 2.0 (P<0.01) on the two external cohorts that include patient outcomes. This artificial intelligence-driven solution offers a rapid, cost-effective, and scalable alternative to genomic testing, with the potential to enhance personalized treatment planning, especially in resource-constrained settings.
BibTeX
@misc{cohen2025prediction,
  author = {Shachar Cohen and Gil Shamai and Edmond Sabo and Alexandra Cretu and Iris Barshack and Tal Goldman and Gil Bar‐Sela and Alexander T. Pearson and Dezheng Huo and Frederick M. Howard and Ron Kimmel and Chen Mayer},
  title = {{Prediction of OncotypeDX recurrence score using H\&E stained WSI images}},
  year = {2025},
  doi = {10.1101/2025.07.21.25331907},
  howpublished = {medRxiv},
}
Figure