AI Test Beats Genomic Tests, Predicts Breast Cancer Return in Hours

AI-driven assay offers faster, cheaper prediction of breast‑cancer relapse

Scientists have unveiled a machine‑learning tool that can forecast the return of breast cancer using routine pathology slides and basic clinical information, cutting the turnaround time and cost compared with existing genomic tests.

Current genomic assays, which assess recurrence risk for hormone‑receptor‑positive tumors, often require weeks to deliver results and consume tissue that cannot be reused. The new approach sidesteps these limitations by leveraging digital images of tumor sections already examined by pathologists.

“Breast cancer comprises diverse subtypes, and choosing an appropriate treatment intensity is challenging,” said Krzysztof J. Geras, a visiting scholar at NYU’s Center for Data Science and adjunct assistant professor at the NYU Grossman School of Medicine, who led the study.

Geras added that the algorithm integrates slide imagery with factors such as tumor stage, patient age, and hormone‑receptor status to generate a personalized risk estimate.

Yann LeCun, professor of computer science and data science at NYU and co‑author of the paper, emphasized that the model’s strength stems from self‑supervised pretraining, which builds rich internal representations before fine‑tuning for the recurrence task.

The research team evaluated the system on data from more than 3,500 individuals drawn from 15 cohorts across seven countries. Performance metrics—including the concordance index and hazard ratio—showed that the AI test consistently distinguished patients at higher versus lower risk of relapse.

Notably, the tool delivered reliable predictions for triple‑negative and HER2‑positive cancers, subtypes that lack robust genomic assays today.

“In testing on thousands of patients, our AI test matched or outperformed a widely used genomic test,” Geras noted. He also highlighted that because the method relies on existing slides, results could be available within hours, reducing expenses and preserving tissue for future analyses.

While the findings are promising, the authors stress the importance of validation in completed randomized clinical trials before the assay can be adopted for routine risk assessment and treatment planning.

Editor’s Note: Some of the paper’s authors hold equity in Ataraxis AI, and NYU retains financial and intellectual‑property interests in the company.