Breast cancer is one of the most common cancers worldwide, and early treatment is key to improving survival rates. However, predicting the risk of cancer recurrence remains a significant challenge. Traditional tools like genomic assays can help, but they have limitations, especially when it comes to different types of breast cancer. A recent breakthrough study introduces a new artificial intelligence (AI) tool that aims to close this gap, offering more accurate and broadly applicable risk assessments for breast cancer patients.
Key facts
- Traditional genomic tests like Oncotype DX are limited, especially for subtypes like triple-negative breast cancer (TNBC) or HER2-positive cases.
- The AI test uses a “vision transformer” to analyze both pathology slides and clinical data, such as ER status, patient age, and nodal involvement, to create a comprehensive risk score.
- Evaluated on over 8,000 breast cancer patients from seven countries, the AI tool outperformed the Oncotype DX test, particularly for TNBC and HER2-positive cancers.
The research team, led by Jan Witowski from Ataraxis AI, NYU Langone Health and NYU Grossman School of Medicine, developed a multi-modal AI test designed to predict the likelihood of breast cancer returning after treatment. This new test integrates both clinical data (such as tumor size and receptor status) and detailed digital pathology images from biopsy samples. Unlike earlier methods that relied on specific gene expression, this approach leverages advanced AI to learn directly from the combination of different types of patient information—making it both unique and powerful.
Tackling an important problem
The current standard of care for breast cancer prognosis often includes genomic tests like Oncotype DX, which use the expression of certain genes to predict the risk of cancer coming back. While useful, these tests have limitations, particularly for certain breast cancer subtypes such as triple-negative breast cancer (TNBC) or HER2-positive cases. Moreover, these tests require special tissue processing and can be costly, which limits accessibility—especially in resource-constrained settings.
To address these issues, the team designed a test that could analyze a broader range of breast cancer types with greater accuracy while using more readily available data. The key idea was to analyze digitized biopsy images using a type of machine learning known as “self-supervised learning”. We trained this AI model, Kestrel, on millions of pathology images from various cancers, enabling it to identify crucial features that human pathologists might miss.
How the AI Tool Works
The multi-modal AI test works by analyzing both pathology slides and clinical information to create a risk score for each patient. It uses a “vision transformer” model, similar to those that power advanced image recognition systems, to extract meaningful patterns from high-resolution biopsy images. We train the model not just to look at a specific set of features, but also to independently learn which features are most predictive of cancer recurrence. We then combine these image-based features with clinical data, including the tumor’s estrogen receptor (ER) status, the patient’s age, and nodal involvement, to create a comprehensive risk score.
We evaluated this new AI test on a dataset of over 8,000 breast cancer patients from seven different countries. It was very accurate; the results showed that it did better than the standard Oncotype DX test, especially at predicting recurrence in harder types of breast cancer, like triple-negative and HER2-positive cancers. The AI tool had a C-index (a measure of how accurate a prediction is) of 0.67 compared to 0.61 for Oncotype DX in a direct comparison of over 850 patients. This suggests that the AI model is a more accurate way to estimate the risk of recurrence.
The real impact for patients and doctors
“This new AI model could significantly change how we approach breast cancer treatment,” said Krzysztof Geras, one of the senior authors of the study. “Our goal was to develop a tool that could accurately predict outcomes across a diverse set of patients, including those for whom existing genomic tests are less effective.”
The AI test holds great promise in enhancing decision-making for patients who may otherwise find themselves in the “gray zone” of risk prediction. For instance, current genomic tests often classify a large proportion of patients as being at an intermediate risk of recurrence, leaving doctors and patients uncertain about whether to pursue aggressive treatment options like chemotherapy. The new AI test, however, reclassified many of these patients into clearer low- or high-risk groups, providing more actionable guidance.
The implications of this research are far-reaching. The AI model could potentially reduce the need for expensive genomic tests, making high-quality cancer prognostication more accessible. Since it uses standard pathology slides that are already part of routine diagnostic procedures, the turnaround time for risk assessment could be much faster, helping doctors make timely decisions about treatment.
With further validation, the researchers hope to integrate this AI model into clinical practice, enabling oncologists to offer more personalized treatment strategies. This also creates opportunities for the application of similar AI models to other types of cancers and medical decision-making.
Witowski stated, “Our ultimate vision is to develop tools that can deploy globally to reduce disparities in cancer care.” “In regions with limited access to specialized genomic testing, this model could be particularly transformative.”
Takeaway
The new multi-modal AI tool for breast cancer prognosis integrates clinical data and digital pathology images, offering a more accurate assessment of recurrence risk compared to traditional genomic tests like Oncotype DX. Developed by a team led by Jan Witowski, the AI model aims to improve decision-making for challenging breast cancer subtypes, particularly triple-negative and HER2-positive cases, while being accessible and reducing the need for expensive testing. By providing more precise risk assessments, the tool could guide more personalized treatment strategies, especially in regions lacking access to specialized tests. This AI-driven approach could become a vital part of the oncologist’s toolkit, transforming breast cancer care globally.
