Breast cancer is the most common cancer in women in the US, with each woman having a 1 in 8 chance of developing breast cancer. Breast cancer is also the second leading cause of cancer death in women, with ~43,000 deaths occurring annually in the US. Treatment and prognosis have been determined by molecular subtype classification, such as expression of HER2, HR, estrogen, and progesterone and more recently the Oncotype DX (ODX) score for predicting recurrence and prognosis. Current assays to test for this recurrence risk, like the Oncotype DX test, have improved personalized therapeutic decision-making, however the time to get results can be up to several weeks and challenges in clinical adoption have occurred due to cost-effectiveness and availability.

Recent advances in artificial intelligence have significantly impacted diagnostic practices. Multiple Instance Learning (MIL) has shown promise for pathology whole slide image analysis and classification. While MIL methodologies and other approaches have been used to predict ODX score, limitations have been seen due to small sample size, reducing the accuracy and clinical applicability of these models.

This invention, AnchorMIL, is a novel deep learning framework that predicts Oncotype DX breast cancer recurrence risk scores directly from digital pathology images, providing both continuous risk scores and binary risk classifications. Using an innovative anchored regression approach, the technology delivers higher accuracy, sensitivity, and specificity than existing MIL-based methods. AnchorMIL presents as a faster, cost-effective alternative to genomic assays that can support personalized treatment decisions while reducing testing time from weeks to hours.

The university is seeking licensing partners and research sponsors to support the continued development and commercialization of AnchorMIL. Companies developing AI-enabled pathology solutions, digital diagnostics, or cancer detection technologies are are encouraged to engage with the university to discuss licensing opportunities, sponsored research agreements, and strategic development partnerships.

Technology Overview

Current Challenges

Current methods for predicting breast cancer recurrence risk, such as the Oncotype DX assay, can be costly, require specialized testing, and often take weeks to return results, delaying treatment decisions. While artificial intelligence-based approaches have been explored to predict recurrence risk directly from pathology images, many existing models are limited by small datasets and insufficient accuracy, reducing their clinical utility and adoption.

Our Innovation

This invention introduces a novel deep learning framework designed to predict ODX risk scores from whole slide images. By using an anchored regression scheme, this model can yield highly accurate continuous risk score predictions, improving both interpretability and performance. Finally, the invention jointly optimizes regression and classification objectives, which allows it to deliver both the continuous score predictions and reliable binary risk stratification, allowing it to be adaptable to a range of clinical decision-making scenarios.

Although other MIL-based methods have been used to determine high/low risk classification, this invention, the AnchorMIL, notably improved accuracy, sensitivity and specificity over several of these other MIL-based methods and has achieved superior performance. This is a promising invention, that is the first to use an anchored regression scheme for prediction of risk and could be a faster and more cost-effective therapeutic strategy compared to current clinically used tests.

Benefits of this Technology

  • Predicts both continuous and binary recurrence risks from digitized histopathology slides.
  • Is cost-effective compared to current genomic assays.
  • Can determine risk scores in just hours, compared to current genomic assays that take weeks to obtain results.
  • Validated on both TCGA-BRCA dataset and OSU in-house dataset, demonstrating the framework’s robustness under real-world clinical variability.
  • Exists as software, that could be easily accessible to multiple clinics.

Stage of Development

AnchorMIL has been fully implemented in PyTorch, an open-source machine learning framework primarily used for developing and training deep learning models, and validated using both The Cancer Genome Atlas and Ohio State University with strong predictive performance. Having achieved proof-of-concept and retrospective validation, the technology is well positioned for prospective clinical evaluation and further translational development.